Re: KAIO Question [8763]

Art, I agree, and having checked for the env variable "KAIO_OFF=1", made sure it was removed from the .profile, and confirmed I do not see it set anywhere else. I have bounced the engine, and confirmed it does not exist, but still do not see the KIO thread. Do I need to do something to turn it back on (i.e., KAIOON=1000 or KAIOOFF=0)? -John -----Original Message----- >From: "ART KAGEL, BLOOMBERG/ 731 LEXIN" <kagel@bloomberg.net> >Sent: Mar 29, 2007 4:54 PM >To: ids@iiug.org >Subject: Re: KAIO Question [8763] > >IDS AIO VPs do NOT use OS Asynch IO facilities, they simulate asynch IO for >cooked chunks. If you use RAW chunks there should be only a few aio vps >configures and onstat -g iov will show one kio thread for each CPU VP. The >only >way you would not see KIO threads if all of your chunks are RAW is if you >disabled KAIO using the environment variable KAIO_OFF=1 in the server's >startup >environment. > >Art S. Kagel >----- Original Message ----- >From: John <ids@iiug.org> >At: 3/29 17:45:04 > >To any AIX / IDS gurus out there, > >Background: > >We are running IDS 10.00 FC4 on AIX 5.2 TL8. We have been having some AIO >issues with two boxes (each is a 390 with 8 cpu's), one dev and one prod. We >tuned the DEV box with min and max AIX AIO 5, and 180 respectively, and it is >running fine, and seemingly wide open. There are actually over 360 IO threads >running, but only a dozen or so showing that they are clocking any CPU time. >This machine, with (all cpu and aio) noage activated, is seemingly screaming >along. >PROD is set with min max 1 and 180 respectively (the machine notes show that >it should be set to MIN =1) and it only has one thread running, and it does >not even show it clocking any CPU. Although on the database side there are 140 >aio vp's (threads), it appears that something is holding AIX from spawning any >more (Unix)AIO processes, and it is seemingly creating a bottleneck. This >database also has noage activated on all cpu and aio vp's, but the machine >seems to not be spawning off any of the seemingly needed (Unix)AIO procs. The >tech does not want to increase the prod config based on the machine notes. >The boxes, prod and dev, have 8 and 128, and 4 and 128 cpu and aio vp's >configured, respectively, again, all noage. Dev is a shared box and only has 4 >cpuvp's. >Tech has showed that although I/O should be in the 80mb per second, we are >using but a small fraction of that. > >Questions: >A. Does setting the AIX MIN aio value to 1 disable the spawning of additional >threads/processes as needed?, and does setting it at say "5" and bouncing the >machine enable and open it up to AIO thread volume increases as AIX deems >appropriate? According to the aforementioned scenario, it certainly appears >to, but if I could get an explanation of exactly why/how it works...??? > >B. On AIX, in regards to the PSE driver, the machine notes for AIX 5.2 and 5.3 >say it should be loaded via "/usr/bin/strload" (generic configuration) and >that it works hand in hand with enabling KAIO. Is there a preferred way as a >DBA to ensure that it is activated on both of our servers and configured >properly? I have read all the documentation I can find, but it does not answer >these specific questions. Does the NETTYPE have to reflect socstr in order to >work? > >C. We have been advised through documentation available on the IBM website >that we should enable KAIO, but have been unable to get it to come on/with the >cooked files, even when building the DB in raw, and have been unable to find >the necessary settings to config to get it to activate. > >I have requested to turn KAIO on for this server. They have cooked >filesystems, and yes, I would alwys use KAIO with raw, but there has been >documentation that shows that you can still use it with KAIO enabled and get >almost raw performance. > >Here is an excerpt from the onconfig file. You can see we have discontinued >using processor affinity until we can get the I/O issue resolved, and scaled >the aiovp's way back. We have reduced the number of aiovps until we can get >the ones that are present to work. This is a data warehouse database with >several fragmented tables, of about 700GB in size. > >....... >SERVERNUM 1 # Unique id corresponding to a OnLine instance >DBSERVERNAME nmprdbdev_shm # Name of default database server>DBSERVERALIASES nmprdbdev_tcp,nmprdbdev_tcp2 # List of alternate dbservernames >SINGLE_CPU_VP 0 # Set to 1 if VPCLASS cpu, num=1 >MULTIPROCESSOR 1 # Set to 1 if VPCLASS cpu, num>1> >#VPCLASS cpu,num=4,aff=4-7,noage # Number of cpu virtual processors (vps) >#VPCLASS aio,num=8,aff=0-7,noage # Number of IO vps > >VPCLASS cpu,num=8 # Number of cpu virtual processors (vps) >VPCLASS aio,num=8 # Number of IO vps (at least 1 per XPSdbspace > >NETTYPE ipcshm,1,250,CPU # Poll thread type,no. of cpu's,max users,class >NETTYPE soctcp,4,250,NET # Poll thread type,no. of cpu's,max users,class >........> >Thanks, > >John > > >******************************************************************************* >Forum Note: Use "Reply" to post a response in the discussion forum. > > >******************************************************************************* > Forum Note: Use "Reply" to post a response in the discussion forum. >