AIX

Hi,

Thanks for your comments once again. I wonder if in this case everytime when we hit the peak in my case we have only 120 AIO servers procesess on this server then any other requests has to sit in the queue and wait for a AIO server to be released/available. If this how it works?. I must admit I'm a bit confuse with the maxreqs parameter. In our enviroment is set to 4096 but I believe this is just a hard limit assuming that the combination of maxservers and the number of CPU's doesn't go beyond this limit.

Code:

│ Asynchronous-I/O-Processes ───────────────────────────────────────────────────────────────────────────────────────────────│
│ Total AIO processes= 120 Actually in use=   0  CPU used=   0.0%    [Use zero                                              │
│       All time peak= 120     Recent peak= 119      Peak=  28.0%     to reset]                                             │
│ Top-Processes-(308) ─────Mode=3  [1=Basic 2=CPU 3=Perf 4=Size 5=I/O 6=Cmds]─────────────

maxservers is the hard limit for the maximum number of aioservers being started, that handle aio requests. They are calculated by the number of logical CPUs * maxservers. They are the workers that grab their work out of the pot with the label "maxreqs".

maxreqs defines how much work may reside in this pot. The single working tasks can be waiting to be handled or already be taken care of by a aioserver which is the asynchronism.
If there is more tasks in there, that are waiting to be taken care of than aioservers, it will just take some time until they are next. If there is far less aioservers than tasks (requests) that are ready to run, then it will be maybe a noticeable delay in processing.

pSeries and AIX Information Center

Hi Again,

Thanks all for you valuable comments. Now we are evaluating either increasing these parameters or going even further and upgrade AIX to 6.1 where I believe AIO is allocated all dynamically. We have a Volume test environment but doesn't have the same amount of resources as in prod so ETL team would not agree to execute a performance and volume test. Thanks Again!!!.

with a large system like this I would encourage you to open a PMR with IBM. there might be other performance tweaking options or design issues.