Multiprocessor setting

Posted in 2010

Summary

Kate asked whether John Miller's old advice still holds: set MULTIPROCESSOR to 0 on 2-CPU boxes to avoid wasted spinning, enabling it only with 3+ CPUs, and whether that applies to IDS 11.5 on quad-core chips. No official answer from Miller appears, but the consensus from Art Kagel, Dick Snoke and others is to count cores rather than sockets, so a quad-core box should use MULTIPROCESSOR 1 (needed for the extra latching when several CPU VPs run concurrently); only a truly single dual-core/single-VP setup might justify 0. Related tips: NUMCPUVPS around cores-1, don't count hyperthreads as cores, and no longer much gain from VPs > cores on Intel. No definitive resolution recorded.

Auto-generated by DrWatson from the posts below — may be imperfect; read the full thread.

Topics: Performance & Tuning, Versions, Editions & End-of-Life

kate — Thu, 22 Jul 2010 14:59:03 — source: IIUG Forums & Mailing Lists

Hey Gurus, A few years back in a performance presentation from John Miller, he discussed the correct tuning for 2 CPU systems. He mentioned that the MULTIPROCESSOR in the config file should actually be set to 0 (off) for two CPU systems to stop the engine from doing unnessary spinning waiting for the next processor. He said that this value only improved the performance of a 3 CPU system or more. My question is - is that still true for version 11.5 and todays new quad core CPUs? Thanks! Kate Tomchik [ kate@iiug.org ] www.iiug.org International Informix Users Group Board of Directors A computer lets you make more mistakes faster than any invention in human history - with the possible exceptions of handguns and tequila. Mitch Ratliffe

↪ replying to kate

Obnoxio The Clown — Thu, 22 Jul 2010 15:01:11 — source: IIUG Forums & Mailing Lists

kate wrote: > Hey Gurus, > A few years back in a performance presentation from John Miller, he > discussed the correct tuning for 2 CPU systems. He mentioned that the > MULTIPROCESSOR in the config file should actually be set to 0 (off) for two > CPU systems to stop the engine from doing unnessary spinning waiting for the > next processor. He said that this value only improved the performance of a > 3 CPU system or more. > > My question is - is that still true for version 11.5 and todays new quad > core CPUs? Personally, I'm not even sure it was true back then. :o) -- Cheers, Obnoxio The Clown http://obotheclown.blogspot.com I will now proceed to pleasure myself with this fish. -- This message has been scanned for viruses and dangerous content by OpenProtect(http://www.openprotect.com), and is believed to be clean.

↪ replying to kate

Wyza, Jonathon — Thu, 22 Jul 2010 15:06:39 — source: IIUG Forums & Mailing Lists

A quad core chip = 4 CPU's. Thus in such an environment MULTIPROCESSOR should be set to 1 (assuming you want informix to use more than just 1 core) Jonathon Wyza CX & CBORD System Administrator CX Programmer/Analyst Administrative Computing Bethel College (574)-257-3381 AIM: Iamwyza jonathon.wyza@bethelcollege.edu ============================== SLES 11x64 & IDS 11.50.FC6 "Don't document the problem, fix it." - Atli Björgvin Oddsson -----Original Message----- From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of kate Sent: Thursday, July 22, 2010 10:59 AM To: ids@iiug.org Subject: Multiprocessor setting [20614] Hey Gurus, A few years back in a performance presentation from John Miller, he discussed the correct tuning for 2 CPU systems. He mentioned that the MULTIPROCESSOR in the config file should actually be set to 0 (off) for two CPU systems to stop the engine from doing unnessary spinning waiting for the next processor. He said that this value only improved the performance of a 3 CPU system or more. My question is - is that still true for version 11.5 and todays new quad core CPUs? Thanks! Kate Tomchik [ kate@iiug.org ] www.iiug.org International Informix Users Group Board of Directors A computer lets you make more mistakes faster than any invention in human history - with the possible exceptions of handguns and tequila. Mitch Ratliffe ******************************************************************************* Forum Note: Use "Reply" to post a response in the discussion forum.

↪ replying to Obnoxio The Clown

Fernando Nunes — Thu, 22 Jul 2010 15:07:16 — source: IIUG Forums & Mailing Lists

I'd say I saw a more recent presentation from John saying that setting it to 0 in a multi-processor machine could give you an impact of 30%... But I maybe wrong... in any case the best thing is to wait for his input... He should be finishing some presentation on OAT, AUS or maybe just running an UPDATE STATS ;) Regards. On Thu, Jul 22, 2010 at 4:01 PM, Obnoxio The Clown <obnoxio@serendipita.com>wrote: > kate wrote: > > Hey Gurus, > > A few years back in a performance presentation from John Miller, he > > discussed the correct tuning for 2 CPU systems. He mentioned that the > > MULTIPROCESSOR in the config file should actually be set to 0 (off) for > two > > CPU systems to stop the engine from doing unnessary spinning waiting for > the > > next processor. He said that this value only improved the performance of > a > > 3 CPU system or more. > > > > My question is - is that still true for version 11.5 and todays new quad > > core CPUs? > > Personally, I'm not even sure it was true back then. :o) > > -- > Cheers, > Obnoxio The Clown > > http://obotheclown.blogspot.com > I will now proceed to pleasure myself with this fish. > > -- > This message has been scanned for viruses and > dangerous content by OpenProtect(http://www.openprotect.com), and is > believed to be clean. > > > > ******************************************************************************* > Forum Note: Use "Reply" to post a response in the discussion forum. > > -- Fernando Nunes Portugal http://informix-technology.blogspot.com My email works... but I don't check it frequently... --001485f875d2c83017048bfb4554

↪ replying to Obnoxio The Clown

Alexandre Marini — Thu, 22 Jul 2010 15:15:46 — source: IIUG Forums & Mailing Lists

Em 22/07/2010 11:01, Obnoxio The Clown escreveu: > kate wrote: >> Hey Gurus, >> A few years back in a performance presentation from John Miller, he >> discussed the correct tuning for 2 CPU systems. He mentioned that the >> MULTIPROCESSOR in the config file should actually be set to 0 (off) for two >> CPU systems to stop the engine from doing unnessary spinning waiting for the >> next processor. He said that this value only improved the performance of a >> 3 CPU system or more. >> >> My question is - is that still true for version 11.5 and todays new quad >> core CPUs? > Personally, I'm not even sure it was true back then. :o) > Hi Kate, here we have several dual-core machines, and all them running 11.50.FC6 on linux_64 archs. All of them are using MULTIPROCESSOR 0, mainly because of low user concurrently sessions, and I think it´s always a "best practice" to let at least 1 cpu core, in order to the operational system needs. On a multi-core cpu, the recommended setting is always 1, with cpu vps affinity, if possible, adjusted to (CORES - 1) calculation. We also have our main servers (multi-cores) using that other settings, just fine (11.50.FC6 also, same arch). Hope it helps. Best regards. -- Alexandre Marini Tecnologia da Informação - DBA msn: alexandre_marini@hotmail.com SEFAZ-MS / SGI-UIMP / Sistemas IBM-Informix Cert-Info-Mgmt_color <Cert-Info-Mgmt_color.jpg> IBM Informix Dynamic Server Certified Professional V10 / V11

↪ replying to kate

Art Kagel — Thu, 22 Jul 2010 16:59:13 — source: IIUG Forums & Mailing Lists

I think that a quad core processor qualifies as a 3 CPU system. At the time that John said that, #cores == #processors. Since that's no longer true, I would go with the number of cores rather than the number of processors. That said, I was always told that if you have more than one core you need to set MULTIPROCESSOR to 1 to add additional locking/latching of memory to prevent simultaneous access to memory structures by more than one CPU VP. John? Art Art S. Kagel Advanced DataTools (www.advancedatatools.com) IIUG Board of Directors (art@iiug.org) Disclaimer: Please keep in mind that my own opinions are my own opinions and do not reflect on my employer, Advanced DataTools, the IIUG, nor any other organization with which I am associated either explicitly, implicitly, or by inference. Neither do those opinions reflect those of other individuals affiliated with any entity with which I am affiliated nor those of the entities themselves. On Thu, Jul 22, 2010 at 10:59 AM, kate <kate@iiug.org> wrote: > Hey Gurus, > A few years back in a performance presentation from John Miller, he > discussed the correct tuning for 2 CPU systems. He mentioned that the > MULTIPROCESSOR in the config file should actually be set to 0 (off) for two > CPU systems to stop the engine from doing unnessary spinning waiting for > the > next processor. He said that this value only improved the performance of a > 3 CPU system or more. > > My question is - is that still true for version 11.5 and todays new quad > core CPUs? > > Thanks! > Kate Tomchik [ kate@iiug.org ] www.iiug.org > International Informix Users Group Board of Directors > > A computer lets you make more mistakes faster than any invention in human > history - with the possible exceptions of handguns and tequila. > Mitch Ratliffe > > > > ******************************************************************************* > Forum Note: Use "Reply" to post a response in the discussion forum. > > --0016e6d26c9a46bc9c048bfcd63e

↪ replying to Art Kagel

Fernando Nunes — Thu, 22 Jul 2010 17:03:52 — source: IIUG Forums & Mailing Lists

This is a very good point. We should not be looking only from a performance point of view, but also from a security point of view... Which again makes me wonder... If we have more than 1 CPU VP, and we know that we have more than one core, why do we need MULTIPROCESSOR? Informix should be able to detect and set it up automatically... Looking forward for some educated hints on this. Regards. On Thu, Jul 22, 2010 at 5:59 PM, Art Kagel <art.kagel@gmail.com> wrote: > I think that a quad core processor qualifies as a 3 CPU system. At the time > that John said that, #cores == #processors. Since that's no longer true, I > would go with the number of cores rather than the number of processors. > > That said, I was always told that if you have more than one core you need > to > set MULTIPROCESSOR to 1 to add additional locking/latching of memory to > prevent simultaneous access to memory structures by more than one CPU VP. > John? > > Art > > Art S. Kagel > Advanced DataTools (www.advancedatatools.com) > IIUG Board of Directors (art@iiug.org) > > Disclaimer: Please keep in mind that my own opinions are my own opinions > and > do not reflect on my employer, Advanced DataTools, the IIUG, nor any other > organization with which I am associated either explicitly, implicitly, or > by > inference. Neither do those opinions reflect those of other individuals > affiliated with any entity with which I am affiliated nor those of the > entities themselves. > > On Thu, Jul 22, 2010 at 10:59 AM, kate <kate@iiug.org> wrote: > > > Hey Gurus, > > A few years back in a performance presentation from John Miller, he > > discussed the correct tuning for 2 CPU systems. He mentioned that the > > MULTIPROCESSOR in the config file should actually be set to 0 (off) for > two > > CPU systems to stop the engine from doing unnessary spinning waiting for > > the > > next processor. He said that this value only improved the performance of > a > > 3 CPU system or more. > > > > My question is - is that still true for version 11.5 and todays new quad > > core CPUs? > > > > Thanks! > > Kate Tomchik [ kate@iiug.org ] www.iiug.org > > International Informix Users Group Board of Directors > > > > A computer lets you make more mistakes faster than any invention in human > > history - with the possible exceptions of handguns and tequila. > > Mitch Ratliffe > > > > > > > > > > ******************************************************************************* > > Forum Note: Use "Reply" to post a response in the discussion forum. > > > > > > --0016e6d26c9a46bc9c048bfcd63e > > > > ******************************************************************************* > Forum Note: Use "Reply" to post a response in the discussion forum. > > -- Fernando Nunes Portugal http://informix-technology.blogspot.com My email works... but I don't check it frequently... --0016e6d7e012e5a97d048bfce6d2

↪ replying to Art Kagel

Paul Watson — Thu, 22 Jul 2010 17:06:50 — source: IIUG Forums & Mailing Lists

I'd also add that we used to be able run NUMCPUVPS > physical CPUs and see a performance increase. I'm not seeing that at all with the Intel multicores Cheers Paul -----Original Message----- From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of Art Kagel Sent: Thursday, July 22, 2010 11:59 AM To: ids@iiug.org Subject: Re: Multiprocessor setting [20619] I think that a quad core processor qualifies as a 3 CPU system. At the time that John said that, #cores == #processors. Since that's no longer true, I would go with the number of cores rather than the number of processors. That said, I was always told that if you have more than one core you need to set MULTIPROCESSOR to 1 to add additional locking/latching of memory to prevent simultaneous access to memory structures by more than one CPU VP. John? Art Art S. Kagel Advanced DataTools (www.advancedatatools.com) IIUG Board of Directors (art@iiug.org) Disclaimer: Please keep in mind that my own opinions are my own opinions and do not reflect on my employer, Advanced DataTools, the IIUG, nor any other organization with which I am associated either explicitly, implicitly, or by inference. Neither do those opinions reflect those of other individuals affiliated with any entity with which I am affiliated nor those of the entities themselves. On Thu, Jul 22, 2010 at 10:59 AM, kate <kate@iiug.org> wrote: > Hey Gurus, > A few years back in a performance presentation from John Miller, he > discussed the correct tuning for 2 CPU systems. He mentioned that the > MULTIPROCESSOR in the config file should actually be set to 0 (off) for two > CPU systems to stop the engine from doing unnessary spinning waiting for > the > next processor. He said that this value only improved the performance of a > 3 CPU system or more. > > My question is - is that still true for version 11.5 and todays new quad > core CPUs? > > Thanks! > Kate Tomchik [ kate@iiug.org ] www.iiug.org > International Informix Users Group Board of Directors > > A computer lets you make more mistakes faster than any invention in human > history - with the possible exceptions of handguns and tequila. > Mitch Ratliffe > > > > **************************************************************************** *** > Forum Note: Use "Reply" to post a response in the discussion forum. > > --0016e6d26c9a46bc9c048bfcd63e **************************************************************************** *** Forum Note: Use "Reply" to post a response in the discussion forum. _____ avast! Antivirus <http://www.avast.com>; : Outbound message clean. Virus Database (VPS): 100722-0, 07/22/2010 Tested on: 7/22/2010 12:06:41 PM avast! - copyright (c) 1988-2010 ALWIL Software.

↪ replying to Paul Watson

Wyza, Jonathon — Thu, 22 Jul 2010 17:12:03 — source: IIUG Forums & Mailing Lists

At this point you should always set the numcpuvps to cores - 1. In the case of a hyper-threaded system (upper end intel xeon/Corei chips) then the max is the number of cores*2 - 1. Jonathon Wyza CX & CBORD System Administrator CX Programmer/Analyst Administrative Computing Bethel College (574)-257-3381 AIM: Iamwyza jonathon.wyza@bethelcollege.edu ============================== SLES 11x64 & IDS 11.50.FC6 "Don't document the problem, fix it." - Atli Björgvin Oddsson -----Original Message----- From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of Paul Watson Sent: Thursday, July 22, 2010 1:07 PM To: ids@iiug.org Subject: RE: Multiprocessor setting [20621] I'd also add that we used to be able run NUMCPUVPS > physical CPUs and see a performance increase. I'm not seeing that at all with the Intel multicores Cheers Paul -----Original Message----- From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of Art Kagel Sent: Thursday, July 22, 2010 11:59 AM To: ids@iiug.org Subject: Re: Multiprocessor setting [20619] I think that a quad core processor qualifies as a 3 CPU system. At the time that John said that, #cores == #processors. Since that's no longer true, I would go with the number of cores rather than the number of processors. That said, I was always told that if you have more than one core you need to set MULTIPROCESSOR to 1 to add additional locking/latching of memory to prevent simultaneous access to memory structures by more than one CPU VP. John? Art Art S. Kagel Advanced DataTools (www.advancedatatools.com) IIUG Board of Directors (art@iiug.org) Disclaimer: Please keep in mind that my own opinions are my own opinions and do not reflect on my employer, Advanced DataTools, the IIUG, nor any other organization with which I am associated either explicitly, implicitly, or by inference. Neither do those opinions reflect those of other individuals affiliated with any entity with which I am affiliated nor those of the entities themselves. On Thu, Jul 22, 2010 at 10:59 AM, kate <kate@iiug.org> wrote: > Hey Gurus, > A few years back in a performance presentation from John Miller, he > discussed the correct tuning for 2 CPU systems. He mentioned that the > MULTIPROCESSOR in the config file should actually be set to 0 (off) > for two > CPU systems to stop the engine from doing unnessary spinning waiting > for the next processor. He said that this value only improved the > performance of a > 3 CPU system or more. > > My question is - is that still true for version 11.5 and todays new > quad core CPUs? > > Thanks! > Kate Tomchik [ kate@iiug.org ] www.iiug.org International Informix > Users Group Board of Directors > > A computer lets you make more mistakes faster than any invention in > human history - with the possible exceptions of handguns and tequila. > Mitch Ratliffe > > > > **************************************************************************** *** > Forum Note: Use "Reply" to post a response in the discussion forum. > > --0016e6d26c9a46bc9c048bfcd63e **************************************************************************** *** Forum Note: Use "Reply" to post a response in the discussion forum. _____ avast! Antivirus <http://www.avast.com>; : Outbound message clean. Virus Database (VPS): 100722-0, 07/22/2010 Tested on: 7/22/2010 12:06:41 PM avast! - copyright (c) 1988-2010 ALWIL Software. ******************************************************************************* Forum Note: Use "Reply" to post a response in the discussion forum.

↪ replying to Wyza, Jonathon

Art Kagel — Thu, 22 Jul 2010 17:41:44 — source: IIUG Forums & Mailing Lists

I wouldn't count hyperthreads as cores at all myself. I need to repeat my old performance testing and determine if setting #VPS > #Cores still helps. I'll let you all know if I get that done. Art Art S. Kagel Advanced DataTools (www.advancedatatools.com) IIUG Board of Directors (art@iiug.org) Disclaimer: Please keep in mind that my own opinions are my own opinions and do not reflect on my employer, Advanced DataTools, the IIUG, nor any other organization with which I am associated either explicitly, implicitly, or by inference. Neither do those opinions reflect those of other individuals affiliated with any entity with which I am affiliated nor those of the entities themselves. On Thu, Jul 22, 2010 at 1:12 PM, Wyza, Jonathon <wyzaj@bethelcollege.edu>wrote: > At this point you should always set the numcpuvps to cores - 1. In the case > of > a hyper-threaded system (upper end intel xeon/Corei chips) then the max is > the > number of cores*2 - 1. > > Jonathon Wyza > CX & CBORD System Administrator > CX Programmer/Analyst > Administrative Computing > Bethel College > (574)-257-3381 > AIM: Iamwyza > jonathon.wyza@bethelcollege.edu > ============================== > SLES 11x64 & IDS 11.50.FC6 > > "Don't document the problem, fix it." > - Atli Björgvin Oddsson > > -----Original Message----- > From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of Paul > Watson > Sent: Thursday, July 22, 2010 1:07 PM > To: ids@iiug.org > Subject: RE: Multiprocessor setting [20621] > > I'd also add that we used to be able run NUMCPUVPS > physical CPUs and see > a > performance increase. I'm not seeing that at all with the Intel multicores > > Cheers > Paul > > -----Original Message----- > From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of Art > Kagel > Sent: Thursday, July 22, 2010 11:59 AM > To: ids@iiug.org > Subject: Re: Multiprocessor setting [20619] > > I think that a quad core processor qualifies as a 3 CPU system. At the time > that John said that, #cores == #processors. Since that's no longer true, I > would go with the number of cores rather than the number of processors. > > That said, I was always told that if you have more than one core you need > to > set MULTIPROCESSOR to 1 to add additional locking/latching of memory to > prevent simultaneous access to memory structures by more than one CPU VP. > John? > > Art > > Art S. Kagel > Advanced DataTools (www.advancedatatools.com) IIUG Board of Directors > (art@iiug.org) > > Disclaimer: Please keep in mind that my own opinions are my own opinions > and > do not reflect on my employer, Advanced DataTools, the IIUG, nor any other > organization with which I am associated either explicitly, implicitly, or > by > inference. Neither do those opinions reflect those of other individuals > affiliated with any entity with which I am affiliated nor those of the > entities themselves. > > On Thu, Jul 22, 2010 at 10:59 AM, kate <kate@iiug.org> wrote: > > > Hey Gurus, > > A few years back in a performance presentation from John Miller, he > > discussed the correct tuning for 2 CPU systems. He mentioned that the > > MULTIPROCESSOR in the config file should actually be set to 0 (off) > > for > two > > CPU systems to stop the engine from doing unnessary spinning waiting > > for the next processor. He said that this value only improved the > > performance of a > > 3 CPU system or more. > > > > My question is - is that still true for version 11.5 and todays new > > quad core CPUs? > > > > Thanks! > > Kate Tomchik [ kate@iiug.org ] www.iiug.org International Informix > > Users Group Board of Directors > > > > A computer lets you make more mistakes faster than any invention in > > human history - with the possible exceptions of handguns and tequila. > > Mitch Ratliffe > > > > > > > > > > **************************************************************************** > *** > > Forum Note: Use "Reply" to post a response in the discussion forum. > > > > > > --0016e6d26c9a46bc9c048bfcd63e > > > **************************************************************************** > *** > Forum Note: Use "Reply" to post a response in the discussion forum. > > _____ > > avast! Antivirus <http://www.avast.com>; : Outbound message clean. > > Virus Database (VPS): 100722-0, 07/22/2010 Tested on: 7/22/2010 12:06:41 PM > avast! - copyright (c) 1988-2010 ALWIL Software. > > > > ******************************************************************************* > Forum Note: Use "Reply" to post a response in the discussion forum. > > > > ******************************************************************************* > Forum Note: Use "Reply" to post a response in the discussion forum. > > --001636920ddb58e406048bfd6e9d

↪ replying to Paul Watson

Art Kagel — Thu, 22 Jul 2010 17:46:54 — source: IIUG Forums & Mailing Lists

Yeah, Intels were always deficient in multiprocessing compared to the other processors out there. For example, while I've recommended that you can treat each 500MHZ of processor speed as an additional CPU VP on most processors, I use about 750-800MHZ on Intel processors. I find the AMD x86 offerings do better at this. Art Art S. Kagel Advanced DataTools (www.advancedatatools.com) IIUG Board of Directors (art@iiug.org) Disclaimer: Please keep in mind that my own opinions are my own opinions and do not reflect on my employer, Advanced DataTools, the IIUG, nor any other organization with which I am associated either explicitly, implicitly, or by inference. Neither do those opinions reflect those of other individuals affiliated with any entity with which I am affiliated nor those of the entities themselves. On Thu, Jul 22, 2010 at 1:06 PM, Paul Watson <paul@oninit.com> wrote: > I'd also add that we used to be able run NUMCPUVPS > physical CPUs and see > a > performance increase. I'm not seeing that at all with the Intel multicores > > Cheers > Paul > > -----Original Message----- > From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of Art > Kagel > Sent: Thursday, July 22, 2010 11:59 AM > To: ids@iiug.org > Subject: Re: Multiprocessor setting [20619] > > I think that a quad core processor qualifies as a 3 CPU system. At the time > that John said that, #cores == #processors. Since that's no longer true, I > would go with the number of cores rather than the number of processors. > > That said, I was always told that if you have more than one core you need > to > set MULTIPROCESSOR to 1 to add additional locking/latching of memory to > prevent simultaneous access to memory structures by more than one CPU VP. > John? > > Art > > Art S. Kagel > Advanced DataTools (www.advancedatatools.com) > IIUG Board of Directors (art@iiug.org) > > Disclaimer: Please keep in mind that my own opinions are my own opinions > and > do not reflect on my employer, Advanced DataTools, the IIUG, nor any other > organization with which I am associated either explicitly, implicitly, or > by > inference. Neither do those opinions reflect those of other individuals > affiliated with any entity with which I am affiliated nor those of the > entities themselves. > > On Thu, Jul 22, 2010 at 10:59 AM, kate <kate@iiug.org> wrote: > > > Hey Gurus, > > A few years back in a performance presentation from John Miller, he > > discussed the correct tuning for 2 CPU systems. He mentioned that the > > MULTIPROCESSOR in the config file should actually be set to 0 (off) for > two > > CPU systems to stop the engine from doing unnessary spinning waiting for > > the > > next processor. He said that this value only improved the performance of > a > > 3 CPU system or more. > > > > My question is - is that still true for version 11.5 and todays new quad > > core CPUs? > > > > Thanks! > > Kate Tomchik [ kate@iiug.org ] www.iiug.org > > International Informix Users Group Board of Directors > > > > A computer lets you make more mistakes faster than any invention in human > > history - with the possible exceptions of handguns and tequila. > > Mitch Ratliffe > > > > > > > > > > **************************************************************************** > *** > > Forum Note: Use "Reply" to post a response in the discussion forum. > > > > > > --0016e6d26c9a46bc9c048bfcd63e > > > **************************************************************************** > *** > Forum Note: Use "Reply" to post a response in the discussion forum. > > _____ > > avast! Antivirus <http://www.avast.com>; : Outbound message clean. > > Virus Database (VPS): 100722-0, 07/22/2010 > Tested on: 7/22/2010 12:06:41 PM > avast! - copyright (c) 1988-2010 ALWIL Software. > > > > ******************************************************************************* > Forum Note: Use "Reply" to post a response in the discussion forum. > > --0015174bde3601b87a048bfd684a

↪ replying to Art Kagel

Lester Knutsen — Thu, 22 Jul 2010 18:11:47 — source: IIUG Forums & Mailing Lists

Art, We can try this out on our next box when you are here next week;) - Lester On 7/22/10 1:41 PM, Art Kagel wrote: > I wouldn't count hyperthreads as cores at all myself. I need to repeat my > old performance testing and determine if setting #VPS> #Cores still helps. > I'll let you all know if I get that done. > > Art > > Art S. Kagel > Advanced DataTools (www.advancedatatools.com) > IIUG Board of Directors (art@iiug.org) > > Disclaimer: Please keep in mind that my own opinions are my own opinions and > do not reflect on my employer, Advanced DataTools, the IIUG, nor any other > organization with which I am associated either explicitly, implicitly, or by > inference. Neither do those opinions reflect those of other individuals > affiliated with any entity with which I am affiliated nor those of the > entities themselves. > > On Thu, Jul 22, 2010 at 1:12 PM, Wyza, Jonathon > <wyzaj@bethelcollege.edu>wrote: > >> At this point you should always set the numcpuvps to cores - 1. In the case >> of >> a hyper-threaded system (upper end intel xeon/Corei chips) then the max is >> the >> number of cores*2 - 1. >> >> Jonathon Wyza >> CX& CBORD System Administrator >> CX Programmer/Analyst >> Administrative Computing >> Bethel College >> (574)-257-3381 >> AIM: Iamwyza >> jonathon.wyza@bethelcollege.edu >> ============================== >> SLES 11x64& IDS 11.50.FC6 >> >> "Don't document the problem, fix it." >> - Atli Björgvin Oddsson >> >> -----Original Message----- >> From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of Paul >> Watson >> Sent: Thursday, July 22, 2010 1:07 PM >> To: ids@iiug.org >> Subject: RE: Multiprocessor setting [20621] >> >> I'd also add that we used to be able run NUMCPUVPS> physical CPUs and see >> a >> performance increase. I'm not seeing that at all with the Intel multicores >> >> Cheers >> Paul >> >> -----Original Message----- >> From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of Art >> Kagel >> Sent: Thursday, July 22, 2010 11:59 AM >> To: ids@iiug.org >> Subject: Re: Multiprocessor setting [20619] >> >> I think that a quad core processor qualifies as a 3 CPU system. At the time >> that John said that, #cores == #processors. Since that's no longer true, I >> would go with the number of cores rather than the number of processors. >> >> That said, I was always told that if you have more than one core you need >> to >> set MULTIPROCESSOR to 1 to add additional locking/latching of memory to >> prevent simultaneous access to memory structures by more than one CPU VP. >> John? >> >> Art >> >> Art S. Kagel >> Advanced DataTools (www.advancedatatools.com) IIUG Board of Directors >> (art@iiug.org) >> >> Disclaimer: Please keep in mind that my own opinions are my own opinions >> and >> do not reflect on my employer, Advanced DataTools, the IIUG, nor any other >> organization with which I am associated either explicitly, implicitly, or >> by >> inference. Neither do those opinions reflect those of other individuals >> affiliated with any entity with which I am affiliated nor those of the >> entities themselves. >> >> On Thu, Jul 22, 2010 at 10:59 AM, kate<kate@iiug.org> wrote: >> >>> Hey Gurus, >>> A few years back in a performance presentation from John Miller, he >>> discussed the correct tuning for 2 CPU systems. He mentioned that the >>> MULTIPROCESSOR in the config file should actually be set to 0 (off) >>> for >> two >>> CPU systems to stop the engine from doing unnessary spinning waiting >>> for the next processor. He said that this value only improved the >>> performance of a >>> 3 CPU system or more. >>> >>> My question is - is that still true for version 11.5 and todays new >>> quad core CPUs? >>> >>> Thanks! >>> Kate Tomchik [ kate@iiug.org ] www.iiug.org International Informix >>> Users Group Board of Directors >>> >>> A computer lets you make more mistakes faster than any invention in >>> human history - with the possible exceptions of handguns and tequila. >>> Mitch Ratliffe >>> >>> >>> >>> >> >> **************************************************************************** >> *** >>> Forum Note: Use "Reply" to post a response in the discussion forum. >>> >>> >> >> --0016e6d26c9a46bc9c048bfcd63e >> >> >> **************************************************************************** >> *** >> Forum Note: Use "Reply" to post a response in the discussion forum. >> >> _____ >> >> avast! Antivirus<http://www.avast.com>; : Outbound message clean. >> >> Virus Database (VPS): 100722-0, 07/22/2010 Tested on: 7/22/2010 12:06:41 PM >> avast! - copyright (c) 1988-2010 ALWIL Software. >> >> >> >> > ******************************************************************************* >> Forum Note: Use "Reply" to post a response in the discussion forum. >> >> >> >> > ******************************************************************************* >> Forum Note: Use "Reply" to post a response in the discussion forum. >> >> > > --001636920ddb58e406048bfd6e9d > > > ******************************************************************************* > Forum Note: Use "Reply" to post a response in the discussion forum. > > > -- ______________________________________________________________________ Lester Knutsen lester@advancedatatools.com Advanced DataTools Corporation Voice: 703-256-0267 x102 Visit our Web page: http://www.advancedatatools.com ______________________________________________________________________

↪ replying to kate

Richard Snoke — Thu, 22 Jul 2010 18:12:47 — source: IIUG Forums & Mailing Lists

Hi, All this is just my opinion. You may legitimately disagree or not believe me. Deal with it. My recollection is that exactly 2 processors, as the term was understood prior to multi-core and/or multi-threaded chips, presents an interesting problem. If you set MULTIPROCESSOR to zero, then there is no need for any of the multiprocessor serializing or synchronizing code. If you do set MULTIPROCESSOR to 1, then all that extra logic is required. In some cases, the overhead of that extra work was worse than just letting things run on a single processor. You did not gain nearly enough to justify using 2 CPU VPs. However, and I think its a big however, newer systems will only have that problem if you have a single dual-core, single threaded processor. Anything more than that is effectively a symmetric multi-processor system, and MULTIPROCESSOR > 0 shouldn't be a problem. If you want to be really picky, and if you have exactly one CPU VP, then perhaps MULTIPROCESSOR at zero makes sense. In that case, be very careful never to add another CPU VP or you will probably have problems due to lack of synchronization. For current systems, I suggest up to (cores*(threads per core))-(some integer less than pi) CPU VPs. But I don't agree that one should always have as many CPU VPs as the system can manage. Use what your workload requires. Cheers, Dick Snoke IBM ChannelWorks dsnoke@us.ibm.com (404) 487-1595 From: "kate" <kate@iiug.org> To: ids@iiug.org Date: 07/22/2010 10:59 AM Subject: Multiprocessor setting [20614] Sent by: ids-bounces@iiug.org Hey Gurus, A few years back in a performance presentation from John Miller, he discussed the correct tuning for 2 CPU systems. He mentioned that the MULTIPROCESSOR in the config file should actually be set to 0 (off) for two CPU systems to stop the engine from doing unnessary spinning waiting for the next processor. He said that this value only improved the performance of a 3 CPU system or more. My question is - is that still true for version 11.5 and todays new quad core CPUs? Thanks! Kate Tomchik [ kate@iiug.org ] www.iiug.org International Informix Users Group Board of Directors A computer lets you make more mistakes faster than any invention in human history - with the possible exceptions of handguns and tequila. Mitch Ratliffe ******************************************************************************* Forum Note: Use "Reply" to post a response in the discussion forum.

↪ replying to Lester Knutsen

Art Kagel — Thu, 22 Jul 2010 18:35:18 — source: IIUG Forums & Mailing Lists

Ya. Art Art S. Kagel Advanced DataTools (www.advancedatatools.com) IIUG Board of Directors (art@iiug.org) Disclaimer: Please keep in mind that my own opinions are my own opinions and do not reflect on my employer, Advanced DataTools, the IIUG, nor any other organization with which I am associated either explicitly, implicitly, or by inference. Neither do those opinions reflect those of other individuals affiliated with any entity with which I am affiliated nor those of the entities themselves. On Thu, Jul 22, 2010 at 2:11 PM, Lester Knutsen <lester@iiug.org> wrote: > Art, > > We can try this out on our next box when you are here next week;) - Lester > > On 7/22/10 1:41 PM, Art Kagel wrote: > > I wouldn't count hyperthreads as cores at all myself. I need to repeat my > > old performance testing and determine if setting #VPS> #Cores still > helps. > > I'll let you all know if I get that done. > > > > Art > > > > Art S. Kagel > > Advanced DataTools (www.advancedatatools.com) > > IIUG Board of Directors (art@iiug.org) > > > > Disclaimer: Please keep in mind that my own opinions are my own opinions > and > > do not reflect on my employer, Advanced DataTools, the IIUG, nor any > other > > organization with which I am associated either explicitly, implicitly, or > by > > inference. Neither do those opinions reflect those of other individuals > > affiliated with any entity with which I am affiliated nor those of the > > entities themselves. > > > > On Thu, Jul 22, 2010 at 1:12 PM, Wyza, Jonathon > > <wyzaj@bethelcollege.edu>wrote: > > > >> At this point you should always set the numcpuvps to cores - 1. In the > case > >> of > >> a hyper-threaded system (upper end intel xeon/Corei chips) then the max > is > >> the > >> number of cores*2 - 1. > >> > >> Jonathon Wyza > >> CX& CBORD System Administrator > >> CX Programmer/Analyst > >> Administrative Computing > >> Bethel College > >> (574)-257-3381 > >> AIM: Iamwyza > >> jonathon.wyza@bethelcollege.edu > >> ============================== > >> SLES 11x64& IDS 11.50.FC6 > >> > >> "Don't document the problem, fix it." > >> - Atli Björgvin Oddsson > >> > >> -----Original Message----- > >> From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of > Paul > >> Watson > >> Sent: Thursday, July 22, 2010 1:07 PM > >> To: ids@iiug.org > >> Subject: RE: Multiprocessor setting [20621] > >> > >> I'd also add that we used to be able run NUMCPUVPS> physical CPUs and > see > >> a > >> performance increase. I'm not seeing that at all with the Intel > multicores > >> > >> Cheers > >> Paul > >> > >> -----Original Message----- > >> From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of > Art > >> Kagel > >> Sent: Thursday, July 22, 2010 11:59 AM > >> To: ids@iiug.org > >> Subject: Re: Multiprocessor setting [20619] > >> > >> I think that a quad core processor qualifies as a 3 CPU system. At the > time > >> that John said that, #cores == #processors. Since that's no longer true, > I > >> would go with the number of cores rather than the number of processors. > >> > >> That said, I was always told that if you have more than one core you > need > >> to > >> set MULTIPROCESSOR to 1 to add additional locking/latching of memory to > >> prevent simultaneous access to memory structures by more than one CPU > VP. > >> John? > >> > >> Art > >> > >> Art S. Kagel > >> Advanced DataTools (www.advancedatatools.com) IIUG Board of Directors > >> (art@iiug.org) > >> > >> Disclaimer: Please keep in mind that my own opinions are my own opinions > >> and > >> do not reflect on my employer, Advanced DataTools, the IIUG, nor any > other > >> organization with which I am associated either explicitly, implicitly, > or > >> by > >> inference. Neither do those opinions reflect those of other individuals > >> affiliated with any entity with which I am affiliated nor those of the > >> entities themselves. > >> > >> On Thu, Jul 22, 2010 at 10:59 AM, kate<kate@iiug.org> wrote: > >> > >>> Hey Gurus, > >>> A few years back in a performance presentation from John Miller, he > >>> discussed the correct tuning for 2 CPU systems. He mentioned that the > >>> MULTIPROCESSOR in the config file should actually be set to 0 (off) > >>> for > >> two > >>> CPU systems to stop the engine from doing unnessary spinning waiting > >>> for the next processor. He said that this value only improved the > >>> performance of a > >>> 3 CPU system or more. > >>> > >>> My question is - is that still true for version 11.5 and todays new > >>> quad core CPUs? > >>> > >>> Thanks! > >>> Kate Tomchik [ kate@iiug.org ] www.iiug.org International Informix > >>> Users Group Board of Directors > >>> > >>> A computer lets you make more mistakes faster than any invention in > >>> human history - with the possible exceptions of handguns and tequila. > >>> Mitch Ratliffe > >>> > >>> > >>> > >>> > >> > >> > > **************************************************************************** > >> *** > >>> Forum Note: Use "Reply" to post a response in the discussion forum. > >>> > >>> > >> > >> --0016e6d26c9a46bc9c048bfcd63e > >> > >> > >> > > **************************************************************************** > >> *** > >> Forum Note: Use "Reply" to post a response in the discussion forum. > >> > >> _____ > >> > >> avast! Antivirus<http://www.avast.com>; : Outbound message clean. > >> > >> Virus Database (VPS): 100722-0, 07/22/2010 Tested on: 7/22/2010 12:06:41 > PM > >> avast! - copyright (c) 1988-2010 ALWIL Software. > >> > >> > >> > >> > > > > ******************************************************************************* > >> Forum Note: Use "Reply" to post a response in the discussion forum. > >> > >> > >> > >> > > > > ******************************************************************************* > >> Forum Note: Use "Reply" to post a response in the discussion forum. > >> > >> > > > > --001636920ddb58e406048bfd6e9d > > > > > > > > ******************************************************************************* > > Forum Note: Use "Reply" to post a response in the discussion forum. > > > > > > > > -- > ______________________________________________________________________ > Lester Knutsen lester@advancedatatools.com > Advanced DataTools Corporation Voice: 703-256-0267 x102 > Visit our Web page: http://www.advancedatatools.com > ______________________________________________________________________ > > > > ******************************************************************************* > Forum Note: Use "Reply" to post a response in the discussion forum. > > --00148531b6a598d82d048bfe2df6

↪ replying to Richard Snoke

Art Kagel — Thu, 22 Jul 2010 18:46:07 — source: IIUG Forums & Mailing Lists

Dick, I agree on several points. If you have only two CPUs/cores whatever, and a single CPU VP, then you may be OK setting MULTIPROCESSOR to zero and I can see where there might be some performance improvement over a single CPU VP with MULTIPROCESSOR set to 1. I also agree that on current systems it's a moot point since most machines today have more than 2 cores per slot and most have two or more CPU slots filled. I also love your rule of thumb based on pi! ;-) But the point about only configuring as many CPU VPs as you need for your workload is specious. Workloads change. If you don't know how many CPU VPs you can safely configure on a given set of hardware and still see incremental gains in either per session performance or overall throughput of the system, then you can't know what to do when your workload increases as it will inevitably do. So, on one face this part of the discussion should be a purely philosophical discussion to determine where that limit is so that when we come to the other face of the discussion and we are running out of steam, we know whether to add a CPU VP or to order another CPU card or an upgraded server. Art Art S. Kagel Advanced DataTools (www.advancedatatools.com) IIUG Board of Directors (art@iiug.org) Disclaimer: Please keep in mind that my own opinions are my own opinions and do not reflect on my employer, Advanced DataTools, the IIUG, nor any other organization with which I am associated either explicitly, implicitly, or by inference. Neither do those opinions reflect those of other individuals affiliated with any entity with which I am affiliated nor those of the entities themselves. On Thu, Jul 22, 2010 at 2:12 PM, Richard Snoke <dsnoke@us.ibm.com> wrote: > Hi, > > All this is just my opinion. You may legitimately disagree or not believe > me. Deal with it. > > My recollection is that exactly 2 processors, as the term was understood > prior to multi-core and/or multi-threaded chips, presents an interesting > problem. If you set MULTIPROCESSOR to zero, then there is no need for any > of the multiprocessor serializing or synchronizing code. If you do set > MULTIPROCESSOR to 1, then all that extra logic is required. In some > cases, the overhead of that extra work was worse than just letting things > run on a single processor. You did not gain nearly enough to justify > using 2 CPU VPs. > > However, and I think its a big however, newer systems will only have that > problem if you have a single dual-core, single threaded processor. > Anything more than that is effectively a symmetric multi-processor system, > and MULTIPROCESSOR > 0 shouldn't be a problem. If you want to be really > picky, and if you have exactly one CPU VP, then perhaps MULTIPROCESSOR at > zero makes sense. In that case, be very careful never to add another CPU > VP or you will probably have problems due to lack of synchronization. > > For current systems, I suggest up to (cores*(threads per core))-(some > integer less than pi) CPU VPs. But I don't agree that one should always > have as many CPU VPs as the system can manage. Use what your workload > requires. > > Cheers, > Dick Snoke > IBM ChannelWorks > > dsnoke@us.ibm.com > (404) 487-1595 > > From: "kate" <kate@iiug.org> > To: ids@iiug.org > Date: 07/22/2010 10:59 AM > Subject: Multiprocessor setting [20614] > Sent by: ids-bounces@iiug.org > > Hey Gurus, > A few years back in a performance presentation from John Miller, he > discussed the correct tuning for 2 CPU systems. He mentioned that the > MULTIPROCESSOR in the config file should actually be set to 0 (off) for > two > CPU systems to stop the engine from doing unnessary spinning waiting for > the > next processor. He said that this value only improved the performance of a > > 3 CPU system or more. > > My question is - is that still true for version 11.5 and todays new quad > core CPUs? > > Thanks! > Kate Tomchik [ kate@iiug.org ] www.iiug.org > International Informix Users Group Board of Directors > > A computer lets you make more mistakes faster than any invention in human > history - with the possible exceptions of handguns and tequila. > Mitch Ratliffe > > > > ******************************************************************************* > > Forum Note: Use "Reply" to post a response in the discussion forum. > > > > ******************************************************************************* > Forum Note: Use "Reply" to post a response in the discussion forum. > > --0016364ece569a6a6f048bfe54f5

↪ replying to Art Kagel

John Miller iii — Thu, 22 Jul 2010 19:00:09 — source: IIUG Forums & Mailing Lists

Just a few comments: 1. The MULTIPROCESSOR onconfig parameter does not have to do with "IF" synchronization to shared memory is done, but rather the type of latch which is done. If you turn on MULTIPROCESSOR onconfig paramater all you latching calls will spin. This is often faster when you are not short on cpu cylces. You can see this spinning in an onstat -g spi. (lots of output). 2. Yes when I was talking in my old presentations about processors this was really prior to cores. So no you should take this to mean cores, not processors. The reason why it might be faster to have MULTIPROCESSOR off in a two core system is if you are running application and other work, the lack of busy spinning can provided time for other operations to run. 3. As for the number of cpu vps exceeding the number of physical cores (see comments about processor threads below). I generally find this is only an advantage when other things are not tuned correctly. We have added a new metric which has simplified this greatly. This new metric is included now in onstat -g glo. It is the Efficiency column for the CPU vps. This tells informix how well the operating system is scheduling the oninit processes. What is the Eff (or Efficiency) column. It compare the amount of time the Hardware core has given the oninit process to run VERSUS the amount of time Informix threads have been running. The efficiency is made into a percentage. Individual virtual processors: vp pid class usercpu syscpu total Thread Eff 1 1452 cpu 997.25 25.06 1022.31 1103.17 92% total = 1022.31 - time the hardware core has give oninit cpu vp to run Thread = 1103.17 - time Informix had threads running Approx 92% of the time oninit is running when it has work, and only 8% of the time it is not running when it has work. This is a very good number. I have been on system where they barely make 40% efficiency. This mean the oninit process is getting less than half of the time on the hardware processor. If this is the case you really need to check the scheduling, noage, and what else is this processor doing. 4. Processor threads. These generally do not help as much as cores. While implementation of these vary widely by hardware vendor. You need to know that these where created to consume the idle time or stalls by hardware core. While database designers generally do try to mitigate these already with many performance tricks, you will receive benefits from them processor threads, but not near as much as cores. John F. Miller III STSM, Support Architect miller3@us.ibm.com 503-578-5645 IBM Informix Dynamic Server (IDS) ids-bounces@iiug.org wrote on 07/22/2010 10:46:55 AM: > [image removed] > > Re: Multiprocessor setting [20624] > > Art Kagel > > to: > > ids > > 07/22/2010 10:47 AM > > Sent by: > > ids-bounces@iiug.org > > Please respond to ids > > Yeah, Intels were always deficient in multiprocessing compared to the other > processors out there. For example, while I've recommended that you can > treat each 500MHZ of processor speed as an additional CPU VP on most > processors, I use about 750-800MHZ on Intel processors. > > I find the AMD x86 offerings do better at this. > > Art > > Art S. Kagel > Advanced DataTools (www.advancedatatools.com) > IIUG Board of Directors (art@iiug.org) > > Disclaimer: Please keep in mind that my own opinions are my own opinions and > do not reflect on my employer, Advanced DataTools, the IIUG, nor any other > organization with which I am associated either explicitly, implicitly, or by > inference. Neither do those opinions reflect those of other individuals > affiliated with any entity with which I am affiliated nor those of the > entities themselves. > > On Thu, Jul 22, 2010 at 1:06 PM, Paul Watson <paul@oninit.com> wrote: > > > I'd also add that we used to be able run NUMCPUVPS > physical CPUs and see > > a > > performance increase. I'm not seeing that at all with the Intel multicores > > > > Cheers > > Paul > > > > -----Original Message----- > > From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of Art > > Kagel > > Sent: Thursday, July 22, 2010 11:59 AM > > To: ids@iiug.org > > Subject: Re: Multiprocessor setting [20619] > > > > I think that a quad core processor qualifies as a 3 CPU system. Atthe time > > that John said that, #cores == #processors. Since that's no longer true, I > > would go with the number of cores rather than the number of processors. > > > > That said, I was always told that if you have more than one core you need > > to > > set MULTIPROCESSOR to 1 to add additional locking/latching of memory to > > prevent simultaneous access to memory structures by more than one CPU VP. > > John? > > > > Art > > > > Art S. Kagel > > Advanced DataTools (www.advancedatatools.com) > > IIUG Board of Directors (art@iiug.org) > > > > Disclaimer: Please keep in mind that my own opinions are my own opinions > > and > > do not reflect on my employer, Advanced DataTools, the IIUG, nor any other > > organization with which I am associated either explicitly, implicitly, or > > by > > inference. Neither do those opinions reflect those of other individuals > > affiliated with any entity with which I am affiliated nor those of the > > entities themselves. > > > > On Thu, Jul 22, 2010 at 10:59 AM, kate <kate@iiug.org> wrote: > > > > > Hey Gurus, > > > A few years back in a performance presentation from John Miller, he > > > discussed the correct tuning for 2 CPU systems. He mentioned that the > > > MULTIPROCESSOR in the config file should actually be set to 0 (off) for > > two > > > CPU systems to stop the engine from doing unnessary spinning waiting for > > > the > > > next processor. He said that this value only improved the performance of > > a > > > 3 CPU system or more. > > > > > > My question is - is that still true for version 11.5 and todays new quad > > > core CPUs? > > > > > > Thanks! > > > Kate Tomchik [ kate@iiug.org ] www.iiug.org > > > International Informix Users Group Board of Directors > > > > > > A computer lets you make more mistakes faster than any inventionin human > > > history - with the possible exceptions of handguns and tequila. > > > Mitch Ratliffe > > > > > > > > > > > > > > > > > **************************************************************************** > > *** > > > Forum Note: Use "Reply" to post a response in the discussion forum. > > > > > > > > > > --0016e6d26c9a46bc9c048bfcd63e > > > > > > > **************************************************************************** > > *** > > Forum Note: Use "Reply" to post a response in the discussion forum. > > > > _____ > > > > avast

↪ replying to Art Kagel

Richard Snoke — Thu, 22 Jul 2010 19:12:44 — source: IIUG Forums & Mailing Lists

Hi Art, I agree with your comments. My only point about the upper bound is that if you really only need 4-6 CPU VPs and configure 24, that's overkill. I prefer to work by adding VPs until the response times are acceptable, then add a couple extra. I agree, you do need to know when you've gone as far as a particular system can go. No need to continue that particular debate. Cheers, Dick Snoke IBM ChannelWorks dsnoke@us.ibm.com (404) 487-1595 From: "Art Kagel" <art.kagel@gmail.com> To: ids@iiug.org Date: 07/22/2010 02:47 PM Subject: Re: Multiprocessor setting [20629] Sent by: ids-bounces@iiug.org Dick, I agree on several points. If you have only two CPUs/cores whatever, and a single CPU VP, then you may be OK setting MULTIPROCESSOR to zero and I can see where there might be some performance improvement over a single CPU VP with MULTIPROCESSOR set to 1. I also agree that on current systems it's a moot point since most machines today have more than 2 cores per slot and most have two or more CPU slots filled. I also love your rule of thumb based on pi! ;-) But the point about only configuring as many CPU VPs as you need for your workload is specious. Workloads change. If you don't know how many CPU VPs you can safely configure on a given set of hardware and still see incremental gains in either per session performance or overall throughput of the system, then you can't know what to do when your workload increases as it will inevitably do. So, on one face this part of the discussion should be a purely philosophical discussion to determine where that limit is so that when we come to the other face of the discussion and we are running out of steam, we know whether to add a CPU VP or to order another CPU card or an upgraded server. Art Art S. Kagel Advanced DataTools (www.advancedatatools.com) IIUG Board of Directors (art@iiug.org) Disclaimer: Please keep in mind that my own opinions are my own opinions and do not reflect on my employer, Advanced DataTools, the IIUG, nor any other organization with which I am associated either explicitly, implicitly, or by inference. Neither do those opinions reflect those of other individuals affiliated with any entity with which I am affiliated nor those of the entities themselves. On Thu, Jul 22, 2010 at 2:12 PM, Richard Snoke <dsnoke@us.ibm.com> wrote: > Hi, > > All this is just my opinion. You may legitimately disagree or not believe > me. Deal with it. > > My recollection is that exactly 2 processors, as the term was understood > prior to multi-core and/or multi-threaded chips, presents an interesting > problem. If you set MULTIPROCESSOR to zero, then there is no need for any > of the multiprocessor serializing or synchronizing code. If you do set > MULTIPROCESSOR to 1, then all that extra logic is required. In some > cases, the overhead of that extra work was worse than just letting things > run on a single processor. You did not gain nearly enough to justify > using 2 CPU VPs. > > However, and I think its a big however, newer systems will only have that > problem if you have a single dual-core, single threaded processor. > Anything more than that is effectively a symmetric multi-processor system, > and MULTIPROCESSOR > 0 shouldn't be a problem. If you want to be really > picky, and if you have exactly one CPU VP, then perhaps MULTIPROCESSOR at > zero makes sense. In that case, be very careful never to add another CPU > VP or you will probably have problems due to lack of synchronization. > > For current systems, I suggest up to (cores*(threads per core))-(some > integer less than pi) CPU VPs. But I don't agree that one should always > have as many CPU VPs as the system can manage. Use what your workload > requires. > > Cheers, > Dick Snoke > IBM ChannelWorks > > dsnoke@us.ibm.com > (404) 487-1595 > > From: "kate" <kate@iiug.org> > To: ids@iiug.org > Date: 07/22/2010 10:59 AM > Subject: Multiprocessor setting [20614] > Sent by: ids-bounces@iiug.org > > Hey Gurus, > A few years back in a performance presentation from John Miller, he > discussed the correct tuning for 2 CPU systems. He mentioned that the > MULTIPROCESSOR in the config file should actually be set to 0 (off) for > two > CPU systems to stop the engine from doing unnessary spinning waiting for > the > next processor. He said that this value only improved the performance of a > > 3 CPU system or more. > > My question is - is that still true for version 11.5 and todays new quad > core CPUs? > > Thanks! > Kate Tomchik [ kate@iiug.org ] www.iiug.org > International Informix Users Group Board of Directors > > A computer lets you make more mistakes faster than any invention in human > history - with the possible exceptions of handguns and tequila. > Mitch Ratliffe > > > > ******************************************************************************* > > Forum Note: Use "Reply" to post a response in the discussion forum. > > > > ******************************************************************************* > Forum Note: Use "Reply" to post a response in the discussion forum. > > --0016364ece569a6a6f048bfe54f5 ******************************************************************************* Forum Note: Use "Reply" to post a response in the discussion forum.

↪ replying to John Miller iii

Wyza, Jonathon — Thu, 22 Jul 2010 19:12:54 — source: IIUG Forums & Mailing Lists

Hmm, that is very interesting. If possible, I'd love some insight on this then: (keep in mind this data is only for 2 days as our server had to be bounced) Individual virtual processors: vp pid class usercpu syscpu total Thread Eff 1 5641 cpu 3081.92 465.82 3547.74 10262.72 34% 3 5654 cpu 2383.56 375.91 2759.47 8820.97 31% 4 5655 cpu 1536.83 265.24 1802.07 4115.01 43% 5 5656 cpu 523.68 57.73 581.41 2616.73 22% 6 5657 cpu 212.99 24.26 237.25 860.44 27% 7 5658 cpu 148.73 15.23 163.96 466.68 35% 8 5659 cpu 117.42 9.84 127.26 289.41 43% When you say scheduling, do you mean IDS scheduling or system/application scheduling? From onconf MULTIPROCESSOR 1 VPCLASS cpu,num=7,noage These processors are shared with apache, and our application. Jonathon Wyza CX & CBORD System Administrator CX Programmer/Analyst Administrative Computing Bethel College (574)-257-3381 AIM: Iamwyza jonathon.wyza@bethelcollege.edu ============================== SLES 11x64 & IDS 11.50.FC6 "Don't document the problem, fix it." - Atli Björgvin Oddsson -----Original Message----- From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of John Miller iii Sent: Thursday, July 22, 2010 3:00 PM To: ids@iiug.org Subject: Re: Multiprocessor setting [20630] Just a few comments: 1. The MULTIPROCESSOR onconfig parameter does not have to do with "IF" synchronization to shared memory is done, but rather the type of latch which is done. If you turn on MULTIPROCESSOR onconfig paramater all you latching calls will spin. This is often faster when you are not short on cpu cylces. You can see this spinning in an onstat -g spi. (lots of output). 2. Yes when I was talking in my old presentations about processors this was really prior to cores. So no you should take this to mean cores, not processors. The reason why it might be faster to have MULTIPROCESSOR off in a two core system is if you are running application and other work, the lack of busy spinning can provided time for other operations to run. 3. As for the number of cpu vps exceeding the number of physical cores (see comments about processor threads below). I generally find this is only an advantage when other things are not tuned correctly. We have added a new metric which has simplified this greatly. This new metric is included now in onstat -g glo. It is the Efficiency column for the CPU vps. This tells informix how well the operating system is scheduling the oninit processes. What is the Eff (or Efficiency) column. It compare the amount of time the Hardware core has given the oninit process to run VERSUS the amount of time Informix threads have been running. The efficiency is made into a percentage. Individual virtual processors: vp pid class usercpu syscpu total Thread Eff 1 1452 cpu 997.25 25.06 1022.31 1103.17 92% total = 1022.31 - time the hardware core has give oninit cpu vp to run Thread = 1103.17 - time Informix had threads running Approx 92% of the time oninit is running when it has work, and only 8% of the time it is not running when it has work. This is a very good number. I have been on system where they barely make 40% efficiency. This mean the oninit process is getting less than half of the time on the hardware processor. If this is the case you really need to check the scheduling, noage, and what else is this processor doing. 4. Processor threads. These generally do not help as much as cores. While implementation of these vary widely by hardware vendor. You need to know that these where created to consume the idle time or stalls by hardware core. While database designers generally do try to mitigate these already with many performance tricks, you will receive benefits from them processor threads, but not near as much as cores. John F. Miller III STSM, Support Architect miller3@us.ibm.com 503-578-5645 IBM Informix Dynamic Server (IDS) ids-bounces@iiug.org wrote on 07/22/2010 10:46:55 AM: > [image removed] > > Re: Multiprocessor setting [20624] > > Art Kagel > > to: > > ids > > 07/22/2010 10:47 AM > > Sent by: > > ids-bounces@iiug.org > > Please respond to ids > > Yeah, Intels were always deficient in multiprocessing compared to the other > processors out there. For example, while I've recommended that you can > treat each 500MHZ of processor speed as an additional CPU VP on most > processors, I use about 750-800MHZ on Intel processors. > > I find the AMD x86 offerings do better at this. > > Art > > Art S. Kagel > Advanced DataTools (www.advancedatatools.com) IIUG Board of Directors > (art@iiug.org) > > Disclaimer: Please keep in mind that my own opinions are my own > opinions and > do not reflect on my employer, Advanced DataTools, the IIUG, nor any other > organization with which I am associated either explicitly, implicitly, > or by > inference. Neither do those opinions reflect those of other > individuals affiliated with any entity with which I am affiliated nor > those of the entities themselves. > > On Thu, Jul 22, 2010 at 1:06 PM, Paul Watson <paul@oninit.com> wrote: > > > I'd also add that we used to be able run NUMCPUVPS > physical CPUs > > and see > > a > > performance increase. I'm not seeing that at all with the Intel multicores > > > > Cheers > > Paul > > > > -----Original Message----- > > From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf > > Of Art > > Kagel > > Sent: Thursday, July 22, 2010 11:59 AM > > To: ids@iiug.org > > Subject: Re: Multiprocessor setting [20619] > > > > I think that a quad core processor qualifies as a 3 CPU system. > > Atthe time > > that John said that, #cores == #processors. Since that's no longer true, I > > would go with the number of cores rather than the number of processors. > > > > That said, I was always told that if you have more than one core you need > > to > > set MULTIPROCESSOR to 1 to add additional locking/latching of memory > > to > > prevent simultaneous access to memory structures by more than one > > CPU VP. > > John? > > > > Art > > > > Art S. Kagel > > Advanced DataTools (www.advancedatatools.com) IIUG Board of > > Directors (art@iiug.org) > > > > Disclaimer: Please keep in mind that my own opinions are my own opinions > > and > > do not reflect on my employer, Advanced DataTools, the IIUG, nor any other > > organization with which I am associated either explicitly, > > implicitly, or > > by > > inference. Neither do those opinions reflect those of other > > individuals > > affiliated with any entity with which I am affiliated nor those of > > the entities themselves. > > > > On Thu, Jul 22, 2010 at 10:59 AM, kate <kate@iiug.org> wrote: > > > > > Hey Gurus, > > > A few years back in a performance presentation from John Miller, > > > he discussed the correct tuning fo

↪ replying to Wyza, Jonathon

Art Kagel — Thu, 22 Jul 2010 19:56:00 — source: IIUG Forums & Mailing Lists

John's talking about OS scheduling. With your applications and a web server all sharing the same set of processors, I'm not surprised that the VP efficiency is around 35%. If most application and web server tasks entail accessing the database then any action requires at least two processes to dig in and start burning CPU cycles. However, efficiency or 35% in this case doesn't necessarily mean that if you separate the apps and web server to separate machines that things will happen 65% faster. It does mean, though, that as your load increases running IDS, apache, and applications on three dedicated boxes instead of sharing just one will allow you to handle three times the load. But that's not any surprise, is it? Part of John's point, IB, is that the CPU VPs are spinning waiting for latches to be released so that they are less likely to be put to sleep by the OS than if they blocked on the latches. That means that Informix may actually have more than 3x capacity available even though the CPU efficiency shows only that much. On a smaller shared box, with only two CPU cores, this spinning can prevent applications from getting enough processor time and it can be better to set MULTIPROCESSOR to zero to eliminate the spinning and release the processors for non-database work. Art Art S. Kagel Advanced DataTools (www.advancedatatools.com) IIUG Board of Directors (art@iiug.org) Disclaimer: Please keep in mind that my own opinions are my own opinions and do not reflect on my employer, Advanced DataTools, the IIUG, nor any other organization with which I am associated either explicitly, implicitly, or by inference. Neither do those opinions reflect those of other individuals affiliated with any entity with which I am affiliated nor those of the entities themselves. On Thu, Jul 22, 2010 at 3:12 PM, Wyza, Jonathon <wyzaj@bethelcollege.edu>wrote: > Hmm, that is very interesting. If possible, I'd love some insight on this > then: (keep in mind this data is only for 2 days as our server had to be > bounced) > > Individual virtual processors: > vp pid class usercpu syscpu total Thread Eff > 1 5641 cpu 3081.92 465.82 3547.74 10262.72 34% > 3 5654 cpu 2383.56 375.91 2759.47 8820.97 31% > 4 5655 cpu 1536.83 265.24 1802.07 4115.01 43% > 5 5656 cpu 523.68 57.73 581.41 2616.73 22% > 6 5657 cpu 212.99 24.26 237.25 860.44 27% > 7 5658 cpu 148.73 15.23 163.96 466.68 35% > 8 5659 cpu 117.42 9.84 127.26 289.41 43% > > When you say scheduling, do you mean IDS scheduling or system/application > scheduling? > > >From onconf > MULTIPROCESSOR 1 > VPCLASS cpu,num=7,noage> > These processors are shared with apache, and our application. > > Jonathon Wyza > CX & CBORD System Administrator > CX Programmer/Analyst > Administrative Computing > Bethel College > (574)-257-3381 > AIM: Iamwyza > jonathon.wyza@bethelcollege.edu > ============================== > SLES 11x64 & IDS 11.50.FC6 > > "Don't document the problem, fix it." > - Atli Björgvin Oddsson > > -----Original Message----- > From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of John > Miller iii > Sent: Thursday, July 22, 2010 3:00 PM > To: ids@iiug.org > Subject: Re: Multiprocessor setting [20630] > > Just a few comments: > > 1. The MULTIPROCESSOR onconfig parameter does not have to do with "IF" > > synchronization to shared memory is done, but rather the type of latch > > which is done. If you turn on MULTIPROCESSOR onconfig paramater > > all you latching calls will spin. This is often faster when you are not > short > > on cpu cylces. You can see this spinning in an onstat -g spi. (lots of > output). > > 2. Yes when I was talking in my old presentations about processors this was > really > > prior to cores. So no you should take this to mean cores, not processors. > > The reason why it might be faster to have MULTIPROCESSOR off in a two > > core system is if you are running application and other work, the lack of > busy > > spinning can provided time for other operations to run. > > 3. As for the number of cpu vps exceeding the number of physical cores (see > comments > > about processor threads below). I generally find this is only an advantage > when > > other things are not tuned correctly. We have added a new metric which has > > simplified this greatly. This new metric is included now in onstat -g glo. > It > is the Efficiency > > column for the CPU vps. This tells informix how well the operating system > is > > scheduling the oninit processes. > > What is the Eff (or Efficiency) column. It compare the amount of time the > > Hardware core has given the oninit process to run VERSUS the amount of time > > Informix threads have been running. The efficiency is made into a > percentage. > > Individual virtual processors: > vp pid class usercpu syscpu total Thread Eff > 1 1452 cpu 997.25 25.06 1022.31 1103.17 92% > > total = 1022.31 - time the hardware core has give oninit cpu vp to run > > Thread = 1103.17 - time Informix had threads running > > Approx 92% of the time oninit is running when it has work, and only 8% of > the > > time it is not running when it has work. This is a very good number. > > I have been on system where they barely make 40% efficiency. This mean the > oninit process is getting less than half of the time on the hardware > processor. If this is the case you really need to check the scheduling, > noage, > and what else is this processor doing. > > 4. Processor threads. > These generally do not help as much as cores. While implementation of these > vary widely by hardware vendor. You need to know that these where created > to > consume the idle time or stalls by hardware core. While database designers > generally do try to mitigate these already with many performance tricks, > you > will receive benefits from them processor threads, but not near as much as > cores. > > John F. Miller III > STSM, Support Architect > miller3@us.ibm.com > 503-578-5645 > IBM Informix Dynamic Server (IDS) > > ids-bounces@iiug.org wrote on 07/22/2010 10:46:55 AM: > > > [image removed] > > > > Re: Multiprocessor setting [20624] > > > > Art Kagel > > > > to: > > > > ids > > > > 07/22/2010 10:47 AM > > > > Sent by: > > > > ids-bounces@iiug.org > > > > Please respond to ids > > > > Yeah, Intels were always deficient in multiprocessing compared to the > other > > processors out there. For example, while I've recommended that you can > > treat each 500MHZ of processor speed as an additional CPU VP on most > > processors, I use about 750-800MHZ on Intel processors. > > > > I find the AMD x86 offerings do better at this. > > > > Art > > > > Art S. Kagel > > Advanced DataTools (www.advancedatatools.com) IIUG Board of Directors > > (art@iiug.org) > > > > Disclaimer: Please keep in mind that my own opinions are my own > > opinions > and > > do not reflect on my employer, Advanced DataTools, the IIUG, nor

↪ replying to Wyza, Jonathon

Alexandre Marini — Thu, 22 Jul 2010 19:57:43 — source: IIUG Forums & Mailing Lists

Hello! I believe Mr John Miller III is talking about operational system IO schedullers. There´s an interesting explaination about the kinds of Linux I/O schedullers on the following presentation (IIUG Conf 2009, I think): C18_Sonkin_Optimizing_Linux_For_IDS Check the recommendations, and if you need to change the default, and could you post another "-g glo" output after, so that we could get an idea of the benefits.... Obs: here our machines uses CentOS, the default IO sched is cfq. But it´s hard to reboot the production systems, and I can´t plan a restart until some months. Maybe someone could help us and answer if it´s really improving something on the VP´s Eff metric? Best regards! Em 22/07/2010 15:12, Wyza, Jonathon escreveu: > Hmm, that is very interesting. If possible, I'd love some insight on this > then: (keep in mind this data is only for 2 days as our server had to be > bounced) > > Individual virtual processors: > vp pid class usercpu syscpu total Thread Eff > 1 5641 cpu 3081.92 465.82 3547.74 10262.72 34% > 3 5654 cpu 2383.56 375.91 2759.47 8820.97 31% > 4 5655 cpu 1536.83 265.24 1802.07 4115.01 43% > 5 5656 cpu 523.68 57.73 581.41 2616.73 22% > 6 5657 cpu 212.99 24.26 237.25 860.44 27% > 7 5658 cpu 148.73 15.23 163.96 466.68 35% > 8 5659 cpu 117.42 9.84 127.26 289.41 43% > > When you say scheduling, do you mean IDS scheduling or system/application > scheduling? > > > From onconf > MULTIPROCESSOR 1 > VPCLASS cpu,num=7,noage> > These processors are shared with apache, and our application. > > Jonathon Wyza > CX& CBORD System Administrator > CX Programmer/Analyst > Administrative Computing > Bethel College > (574)-257-3381 > AIM: Iamwyza > jonathon.wyza@bethelcollege.edu > ============================== > SLES 11x64& IDS 11.50.FC6 > > "Don't document the problem, fix it." > - Atli Björgvin Oddsson > > -----Original Message----- > From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of John > Miller iii > Sent: Thursday, July 22, 2010 3:00 PM > To: ids@iiug.org > Subject: Re: Multiprocessor setting [20630] > > Just a few comments: > > 1. The MULTIPROCESSOR onconfig parameter does not have to do with "IF" > > synchronization to shared memory is done, but rather the type of latch > > which is done. If you turn on MULTIPROCESSOR onconfig paramater > > all you latching calls will spin. This is often faster when you are not short > > on cpu cylces. You can see this spinning in an onstat -g spi. (lots of > output). > > 2. Yes when I was talking in my old presentations about processors this was > really > > prior to cores. So no you should take this to mean cores, not processors. > > The reason why it might be faster to have MULTIPROCESSOR off in a two > > core system is if you are running application and other work, the lack of busy > > spinning can provided time for other operations to run. > > 3. As for the number of cpu vps exceeding the number of physical cores (see > comments > > about processor threads below). I generally find this is only an advantage > when > > other things are not tuned correctly. We have added a new metric which has > > simplified this greatly. This new metric is included now in onstat -g glo. It > is the Efficiency > > column for the CPU vps. This tells informix how well the operating system is > > scheduling the oninit processes. > > What is the Eff (or Efficiency) column. It compare the amount of time the > > Hardware core has given the oninit process to run VERSUS the amount of time > > Informix threads have been running. The efficiency is made into a percentage. > > Individual virtual processors: > vp pid class usercpu syscpu total Thread Eff > 1 1452 cpu 997.25 25.06 1022.31 1103.17 92% > > total = 1022.31 - time the hardware core has give oninit cpu vp to run > > Thread = 1103.17 - time Informix had threads running > > Approx 92% of the time oninit is running when it has work, and only 8% of the > > time it is not running when it has work. This is a very good number. > > I have been on system where they barely make 40% efficiency. This mean the > oninit process is getting less than half of the time on the hardware > processor. If this is the case you really need to check the scheduling, noage, > and what else is this processor doing. > > 4. Processor threads. > These generally do not help as much as cores. While implementation of these > vary widely by hardware vendor. You need to know that these where created to > consume the idle time or stalls by hardware core. While database designers > generally do try to mitigate these already with many performance tricks, you > will receive benefits from them processor threads, but not near as much as > cores. > > John F. Miller III > STSM, Support Architect > miller3@us.ibm.com > 503-578-5645 > IBM Informix Dynamic Server (IDS) > > ids-bounces@iiug.org wrote on 07/22/2010 10:46:55 AM: > >> [image removed] >> >> Re: Multiprocessor setting [20624] >> >> Art Kagel >> >> to: >> >> ids >> >> 07/22/2010 10:47 AM >> >> Sent by: >> >> ids-bounces@iiug.org >> >> Please respond to ids >> >> Yeah, Intels were always deficient in multiprocessing compared to the > other >> processors out there. For example, while I've recommended that you can >> treat each 500MHZ of processor speed as an additional CPU VP on most >> processors, I use about 750-800MHZ on Intel processors. >> >> I find the AMD x86 offerings do better at this. >> >> Art >> >> Art S. Kagel >> Advanced DataTools (www.advancedatatools.com) IIUG Board of Directors >> (art@iiug.org) >> >> Disclaimer: Please keep in mind that my own opinions are my own >> opinions > and >> do not reflect on my employer, Advanced DataTools, the IIUG, nor any > other >> organization with which I am associated either explicitly, implicitly, >> or > by >> inference. Neither do those opinions reflect those of other >> individuals affiliated with any entity with which I am affiliated nor >> those of the entities themselves. >> >> On Thu, Jul 22, 2010 at 1:06 PM, Paul Watson<paul@oninit.com> wrote: >> >>> I'd also add that we used to be able run NUMCPUVPS> physical CPUs >>> and > see >>> a >>> performance increase. I'm not seeing that at all with the Intel > multicores >>> Cheers >>> Paul >>> >>> -----Original Message----- >>> From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf >>> Of > Art >>> Kagel >>> Sent: Thursday, July 22, 2010 11:59 AM >>> To: ids@iiug.org >>> Subject: Re: Multiprocessor setting [20619] >>> >>> I think that a quad core processor qualifies as a 3 CPU system. >>> Atthe > time >>> that John said that, #cores == #processors. Since that's no longer > true, I >>> would go with the number of cores rather than the number of processors. >>> That said, I was always told that if you have more than one core you@@NL@

↪ replying to John Miller iii

Fernando Nunes — Thu, 22 Jul 2010 22:27:34 — source: IIUG Forums & Mailing Lists

One side note: There was a bug with the efficiency metric IF you run onstat -z. Some counters were not reset and the calculation of Efficiency became 0 (with time) It's fixed in 11.50.xC6. Systems where onstat -z is not run should not be affected. Reference: idsdb00189503 Regards. On Thu, Jul 22, 2010 at 8:00 PM, John Miller iii <miller3@us.ibm.com> wrote: > Just a few comments: > > 1. The MULTIPROCESSOR onconfig parameter does not have to do with "IF" > > synchronization to shared memory is done, but rather the type of > latch > > which is done. If you turn on MULTIPROCESSOR onconfig paramater > > all you latching calls will spin. This is often faster when you are > not short > > on cpu cylces. You can see this spinning in an onstat -g spi. (lots > of output). > > 2. Yes when I was talking in my old presentations about processors this > was really > > prior to cores. So no you should take this to mean cores, not > processors. > > The reason why it might be faster to have MULTIPROCESSOR off in a > two > > core system is if you are running application and other work, the > lack of busy > > spinning can provided time for other operations to run. > > 3. As for the number of cpu vps exceeding the number of physical cores > (see comments > > about processor threads below). I generally find this is only an > advantage when > > other things are not tuned correctly. We have added a new metric > which has > > simplified this greatly. This new metric is included now in onstat > -g glo. It is the Efficiency > > column for the CPU vps. This tells informix how well the operating > system is > > scheduling the oninit processes. > > What is the Eff (or Efficiency) column. It compare the amount of > time the > > Hardware core has given the oninit process to run VERSUS the amount > of time > > Informix threads have been running. The efficiency is made into a > percentage. > > Individual virtual processors: > vp pid class usercpu syscpu total Thread Eff > 1 1452 cpu 997.25 25.06 1022.31 1103.17 92% > > total = 1022.31 - time the hardware core has give oninit cpu vp to > run > > Thread = 1103.17 - time Informix had threads running > > Approx 92% of the time oninit is running when it has work, and only > 8% of the > > time it is not running when it has work. This is a very good > number. > > I have been on system where they barely make 40% efficiency. This mean the > oninit process is getting less than half of the time on the hardware > processor. If > this is the case you really need to check the scheduling, noage, and what > else > is this processor doing. > > 4. Processor threads. > These generally do not help as much as cores. While implementation of > these vary > widely by hardware vendor. You need to know that these where created to > consume > the idle time or stalls by hardware core. While database designers > generally do try > to mitigate these already with many performance tricks, you will receive > benefits > from them processor threads, but not near as much as cores. > > John F. Miller III > STSM, Support Architect > miller3@us.ibm.com > 503-578-5645 > IBM Informix Dynamic Server (IDS) > > ids-bounces@iiug.org wrote on 07/22/2010 10:46:55 AM: > > > [image removed] > > > > Re: Multiprocessor setting [20624] > > > > Art Kagel > > > > to: > > > > ids > > > > 07/22/2010 10:47 AM > > > > Sent by: > > > > ids-bounces@iiug.org > > > > Please respond to ids > > > > Yeah, Intels were always deficient in multiprocessing compared to the > other > > processors out there. For example, while I've recommended that you can > > treat each 500MHZ of processor speed as an additional CPU VP on most > > processors, I use about 750-800MHZ on Intel processors. > > > > I find the AMD x86 offerings do better at this. > > > > Art > > > > Art S. Kagel > > Advanced DataTools (www.advancedatatools.com) > > IIUG Board of Directors (art@iiug.org) > > > > Disclaimer: Please keep in mind that my own opinions are my own opinions > and > > do not reflect on my employer, Advanced DataTools, the IIUG, nor any > other > > organization with which I am associated either explicitly, implicitly, or > by > > inference. Neither do those opinions reflect those of other individuals > > affiliated with any entity with which I am affiliated nor those of the > > entities themselves. > > > > On Thu, Jul 22, 2010 at 1:06 PM, Paul Watson <paul@oninit.com> wrote: > > > > > I'd also add that we used to be able run NUMCPUVPS > physical CPUs and > see > > > a > > > performance increase. I'm not seeing that at all with the Intel > multicores > > > > > > Cheers > > > Paul > > > > > > -----Original Message----- > > > From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of > Art > > > Kagel > > > Sent: Thursday, July 22, 2010 11:59 AM > > > To: ids@iiug.org > > > Subject: Re: Multiprocessor setting [20619] > > > > > > I think that a quad core processor qualifies as a 3 CPU system. Atthe > time > > > that John said that, #cores == #processors. Since that's no longer > true, I > > > would go with the number of cores rather than the number of processors. > > > > > > > That said, I was always told that if you have more than one core you > need > > > to > > > set MULTIPROCESSOR to 1 to add additional locking/latching of memory to > > > > prevent simultaneous access to memory structures by more than one CPU > VP. > > > John? > > > > > > Art > > > > > > Art S. Kagel > > > Advanced DataTools (www.advancedatatools.com) > > > IIUG Board of Directors (art@iiug.org) > > > > > > Disclaimer: Please keep in mind that my own opinions are my own > opinions > > > and > > > do not reflect on my employer, Advanced DataTools, the IIUG, nor any > other > > > organization with which I am associated either explicitly, implicitly, > or > > > by > > > inference. Neither do those opinions reflect those of other individuals > > > > affiliated with any entity with which I am affiliated nor those of the > > > entities themselves. > > > > > > On Thu, Jul 22, 2010 at 10:59 AM, kate <kate@iiug.org> wrote: > > > > > > > Hey Gurus, > > > > A few years back in a performance presentation from John Miller, he > > > > discussed the correct tuning for 2 CPU systems. He mentioned that the > > > > > MULTIPROCESSOR in the config file should actually be set to 0 (off) > for > > > two > > > > CPU systems to stop the engine from doing unnessary spinning waiting > for > > > > the > > > > next processor. He said that this value only improved the performance > of > > > a > > > > 3 CPU system or more. > > > > > > > > My question is - is that still true for version 11.5 and todays new > quad > > > > core CPUs? > > > > > > > > Thanks! > > > > Kate Tomchik [ kate@iiug.org ] www.iiug.org > > > > International Informix User

Multiprocessor setting

Related threads

Your privacy choices