cpuvp efficiency AIX 7.1/IDS 11.5.FC8

AIX 7.1 lpar info:mode=Uncapped smt=On lcpu=16 mem=32768MB psize=6 ent=2.00 16 CPUVPS; IDS 11.5.FC8 Question regarding cpu efficiency (a metric I've never played much with). An "onstat -g glo" shows cpuvp efficiency ranges from 12-14% all the time. Similar in production (bigger box, higher entitlement, etc.) I'm not worried about "load balancing" - in earlier IDS we didn't try to load balance/time slice, etc. and haven't heard that that has changed. Anyone have concerns about efficiency being in the this range all the time? Thanks - Mark Scranton The Mark Scranton Group mark@markscranton.com

Hello, Mark. 11.50 engine was not too accurate in onstat -g glo output (eff column). You should consider onstat -g glo (at least) in 11.70 (latest release) or even better in 12.10 engines. Consider an upgrade, mainly because 11.50 will soon enter in EOS. Regards. Alexandre Marini IBM Informix Certified Professional v10 / v11.50 / v11.70 / v12.10 IBM Information Management Informix Technical Professional IBM Certified Developer - Informix Genero BRIUG website administrator Informix independent consultant > To: ids@iiug.org > From: mark@markscranton.com > Subject: cpuvp efficiency AIX 7.1/IDS 11.5.FC8 [35331] > Date: Fri, 26 Jun 2015 09:31:52 -0400 > > AIX 7.1 > lpar info:mode=Uncapped smt=On lcpu=16 mem=32768MB psize=6 ent=2.00 > 16 CPUVPS; IDS 11.5.FC8 > > Question regarding cpu efficiency (a metric I've never played much with). An > "onstat -g glo" shows cpuvp efficiency ranges from 12-14% all the time. > Similar in production (bigger box, higher entitlement, etc.) I'm not worried > about "load balancing" - in earlier IDS we didn't try to load balance/time > slice, etc. and haven't heard that that has changed. Anyone have concerns > about efficiency being in the this range all the time? > > Thanks - > Mark Scranton > The Mark Scranton Group > mark@markscranton.com > > > ******************************************************************************* > Forum Note: Use "Reply" to post a response in the discussion forum. >

Mark: As I understand it, the efficiency is a ratio of time spent by the CPU VPs in processing server threads against total CPU time which would include polling, spinning, etc. In theory the higher the efficiency the fewer cycles are being wasted. Some use it to determine whether there are too many CPU VPs (if efficiency is low). I do not follow that myself. Assuming that there are more concurrent client sessions I want sufficient CPU VPs to process all active user threads (so sqlexec, sort threads, scan threads, merge threads, etc.) in parallel. Or, at least, as many as the CPU cores can handle. When I overclock the CPU VPs (which is what I call it when there are more CPU VPs than processor cores) the efficiency of the CPU VPs goes down so those who follow the efficiency would deduce that there are more VPs than are needed. However, I know that overclocked the server is more responsive and gets more work done. A perfect example is a client I recently worked with. They were running a P7 system with 5 processor cores (so 20 hardware SMP threads) and 5 CPU VPs and processing 20 complex transactions a second. I overclocked them to 20 CPU VPs and increased throughput to 120 transactions per second pushing the cores to 99% utilization. So a 300% increase in CPU VPs resulted in a 500% increase in throughput. We needed more cores to reach their SLA requirements so we tested the theory that adding CPU cores would scale linearly. First we added one core without adding any CPU VPs. Throughput increase by a small amount but not significantly and idle time was around 8-12% across all 6 cores. Then I added four more CPU VPs an 18% increase in the number of VPs. Processor idle time went down to 1-3% and transaction throughput improved by ... 18%, Efficiency with one VP per core was around 75% with 40% idle time on the processors. With four VPs per core it was around 25% with 1-3% idle time. Far worse "efficiency" numbers but far better performance! Do I recommend running the machine at close to 100% under normal conditions? No, they need headroom for growth, but that just means they need more cores to meet their SLAs and also have some headroom on the box for the future. Art Art S. Kagel, President and Principal Consultant ASK Database Management www.askdbmgt.com Blog: http://informix-myview.blogspot.com/ Disclaimer: Please keep in mind that my own opinions are my own opinions and do not reflect on 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 Fri, Jun 26, 2015 at 9:31 AM, MARK SCRANTON <mark@markscranton.com> wrote: > AIX 7.1 > lpar info:mode=Uncapped smt=On lcpu=16 mem=32768MB psize=6 ent=2.00 > 16 CPUVPS; IDS 11.5.FC8 > > Question regarding cpu efficiency (a metric I've never played much with). > An > "onstat -g glo" shows cpuvp efficiency ranges from 12-14% all the time. > Similar in production (bigger box, higher entitlement, etc.) I'm not > worried > about "load balancing" - in earlier IDS we didn't try to load balance/time > slice, etc. and haven't heard that that has changed. Anyone have concerns > about efficiency being in the this range all the time? > > Thanks - > Mark Scranton > The Mark Scranton Group > mark@markscranton.com > > > > ******************************************************************************* > Forum Note: Use "Reply" to post a response in the discussion forum. > > --047d7bdc11bc0bbc6d05196cc796

I would also add, test with SMT off - I have found you can drive the server harder sooner with SMT disabled, although the absolute max throughput was the same. Your sysadmins will hate turning off SMT, but it is an easy test to do Cheers Paul > Mark: > > As I understand it, the efficiency is a ratio of time spent by the CPU VPs > in processing server threads against total CPU time which would include > polling, spinning, etc. In theory the higher the efficiency the fewer > cycles are being wasted. Some use it to determine whether there are too > many CPU VPs (if efficiency is low). I do not follow that myself. > Assuming that there are more concurrent client sessions I want sufficient > CPU VPs to process all active user threads (so sqlexec, sort threads, scan > threads, merge threads, etc.) in parallel. Or, at least, as many as the > CPU cores can handle. > > When I overclock the CPU VPs (which is what I call it when there are more > CPU VPs than processor cores) the efficiency of the CPU VPs goes down so > those who follow the efficiency would deduce that there are more VPs than > are needed. However, I know that overclocked the server is more responsive > and gets more work done. > > A perfect example is a client I recently worked with. They were running a > P7 system with 5 processor cores (so 20 hardware SMP threads) and 5 CPU > VPs > and processing 20 complex transactions a second. I overclocked them to 20 > CPU VPs and increased throughput to 120 transactions per second pushing > the > cores to 99% utilization. So a 300% increase in CPU VPs resulted in a 500% > increase in throughput. We needed more cores to reach their SLA > requirements so we tested the theory that adding CPU cores would scale > linearly. First we added one core without adding any CPU VPs. Throughput > increase by a small amount but not significantly and idle time was around > 8-12% across all 6 cores. Then I added four more CPU VPs an 18% increase > in the number of VPs. Processor idle time went down to 1-3% and > transaction throughput improved by ... 18%, > > Efficiency with one VP per core was around 75% with 40% idle time on the > processors. With four VPs per core it was around 25% with 1-3% idle time. > Far worse "efficiency" numbers but far better performance! > > Do I recommend running the machine at close to 100% under normal > conditions? No, they need headroom for growth, but that just means they > need more cores to meet their SLAs and also have some headroom on the box > for the future. > > Art > > Art S. Kagel, President and Principal Consultant > ASK Database Management > www.askdbmgt.com > > Blog: http://informix-myview.blogspot.com/ > > Disclaimer: Please keep in mind that my own opinions are my own opinions > and do not reflect on 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 Fri, Jun 26, 2015 at 9:31 AM, MARK SCRANTON <mark@markscranton.com> > wrote: > >> AIX 7.1 >> lpar info:mode=Uncapped smt=On lcpu=16 mem=32768MB psize=6 ent=2.00 >> 16 CPUVPS; IDS 11.5.FC8 >> >> Question regarding cpu efficiency (a metric I've never played much >> with). >> An >> "onstat -g glo" shows cpuvp efficiency ranges from 12-14% all the time. >> Similar in production (bigger box, higher entitlement, etc.) I'm not >> worried >> about "load balancing" - in earlier IDS we didn't try to load >> balance/time >> slice, etc. and haven't heard that that has changed. Anyone have >> concerns >> about efficiency being in the this range all the time? >> >> Thanks - >> Mark Scranton >> The Mark Scranton Group >> mark@markscranton.com >> >> >> >> > ******************************************************************************* >> Forum Note: Use "Reply" to post a response in the discussion forum. >> >> > > --047d7bdc11bc0bbc6d05196cc796 > > > ******************************************************************************* > Forum Note: Use "Reply" to post a response in the discussion forum. > -- Paul Watson Tel: +1 913-674-0360 Mob: +1 913-387-7529 Web: www.oninit.com Oninit® is a registered trademark of Oninit LLC Failure is not as frightening as regret. If you want to improve, be content to be thought foolish and stupid. What this country needs are more unemployed politicians

Limiting to two threads instead of 4 also helps. Art Art S. Kagel, President and Principal Consultant ASK Database Management www.askdbmgt.com Blog: http://informix-myview.blogspot.com/ Disclaimer: Please keep in mind that my own opinions are my own opinions and do not reflect on 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 Fri, Jun 26, 2015 at 10:50 AM, Paul Watson <paul@oninit.com> wrote: > I would also add, test with SMT off - I have found you can drive the > server harder sooner with SMT disabled, although the absolute max > throughput was the same. Your sysadmins will hate turning off SMT, but it > is an easy test to do > > Cheers > Paul > > > Mark: > > > > As I understand it, the efficiency is a ratio of time spent by the CPU > VPs > > in processing server threads against total CPU time which would include > > polling, spinning, etc. In theory the higher the efficiency the fewer > > cycles are being wasted. Some use it to determine whether there are too > > many CPU VPs (if efficiency is low). I do not follow that myself. > > Assuming that there are more concurrent client sessions I want sufficient > > CPU VPs to process all active user threads (so sqlexec, sort threads, > scan > > threads, merge threads, etc.) in parallel. Or, at least, as many as the > > CPU cores can handle. > > > > When I overclock the CPU VPs (which is what I call it when there are more > > CPU VPs than processor cores) the efficiency of the CPU VPs goes down so > > those who follow the efficiency would deduce that there are more VPs than > > are needed. However, I know that overclocked the server is more > responsive > > and gets more work done. > > > > A perfect example is a client I recently worked with. They were running a > > P7 system with 5 processor cores (so 20 hardware SMP threads) and 5 CPU > > VPs > > and processing 20 complex transactions a second. I overclocked them to 20 > > CPU VPs and increased throughput to 120 transactions per second pushing > > the > > cores to 99% utilization. So a 300% increase in CPU VPs resulted in a > 500% > > increase in throughput. We needed more cores to reach their SLA > > requirements so we tested the theory that adding CPU cores would scale > > linearly. First we added one core without adding any CPU VPs. Throughput > > increase by a small amount but not significantly and idle time was around > > 8-12% across all 6 cores. Then I added four more CPU VPs an 18% increase > > in the number of VPs. Processor idle time went down to 1-3% and > > transaction throughput improved by ... 18%, > > > > Efficiency with one VP per core was around 75% with 40% idle time on the > > processors. With four VPs per core it was around 25% with 1-3% idle time. > > Far worse "efficiency" numbers but far better performance! > > > > Do I recommend running the machine at close to 100% under normal > > conditions? No, they need headroom for growth, but that just means they > > need more cores to meet their SLAs and also have some headroom on the box > > for the future. > > > > Art > > > > Art S. Kagel, President and Principal Consultant > > ASK Database Management > > www.askdbmgt.com > > > > Blog: http://informix-myview.blogspot.com/ > > > > Disclaimer: Please keep in mind that my own opinions are my own opinions > > and do not reflect on 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 Fri, Jun 26, 2015 at 9:31 AM, MARK SCRANTON <mark@markscranton.com> > > wrote: > > > >> AIX 7.1 > >> lpar info:mode=Uncapped smt=On lcpu=16 mem=32768MB psize=6 ent=2.00 > >> 16 CPUVPS; IDS 11.5.FC8 > >> > >> Question regarding cpu efficiency (a metric I've never played much > >> with). > >> An > >> "onstat -g glo" shows cpuvp efficiency ranges from 12-14% all the time. > >> Similar in production (bigger box, higher entitlement, etc.) I'm not > >> worried > >> about "load balancing" - in earlier IDS we didn't try to load > >> balance/time > >> slice, etc. and haven't heard that that has changed. Anyone have > >> concerns > >> about efficiency being in the this range all the time? > >> > >> Thanks - > >> Mark Scranton > >> The Mark Scranton Group > >> mark@markscranton.com > >> > >> > >> > >> > > > > ******************************************************************************* > >> Forum Note: Use "Reply" to post a response in the discussion forum. > >> > >> > > > > --047d7bdc11bc0bbc6d05196cc796 > > > > > > > > ******************************************************************************* > > Forum Note: Use "Reply" to post a response in the discussion forum. > > > > -- > Paul Watson > Tel: +1 913-674-0360 > Mob: +1 913-387-7529 > Web: www.oninit.com > > Oninit® is a registered trademark of Oninit LLC > > Failure is not as frightening as regret. > If you want to improve, be content to be thought foolish and stupid. > What this country needs are more unemployed politicians > > > > ******************************************************************************* > Forum Note: Use "Reply" to post a response in the discussion forum. > > --bcaec519694121172a05196ceaa8

Is KAIO enabled? If efficiency in onstat -g glo is (total / Thread) * 100, Thread time will increase while waiting for KAIO but total CPU time will not increase leading to "lower efficiency". I've noticed that disabling KAIO can change my onstat -g glo efficiency from 50% to high 90%. I didn't disable KAIO, I don't worry too much about CPU efficiency when KAIO is enabled. Andrew -----Original Message----- From: ids-bounces@iiug.org [mailto:ids-bounces@iiug.org] On Behalf Of MARK SCRANTON Sent: Friday, June 26, 2015 8:32 AM To: ids@iiug.org Subject: cpuvp efficiency AIX 7.1/IDS 11.5.FC8 [35331] AIX 7.1 lpar info:mode=Uncapped smt=On lcpu=16 mem=32768MB psize=6 ent=2.00 16 CPUVPS; IDS 11.5.FC8 Question regarding cpu efficiency (a metric I've never played much with). An "onstat -g glo" shows cpuvp efficiency ranges from 12-14% all the time. Similar in production (bigger box, higher entitlement, etc.) I'm not worried about "load balancing" - in earlier IDS we didn't try to load balance/time slice, etc. and haven't heard that that has changed. Anyone have concerns about efficiency being in the this range all the time? Thanks - Mark Scranton The Mark Scranton Group mark@markscranton.com **************************************************************************** *** Forum Note: Use "Reply" to post a response in the discussion forum.