Am currently running some performance testing in preperation for a migration from 8.1 aix to 8.7 linux.
I am running a very simple test Server job to populate a 64 bit hash file with several million rows. The same job with the same data and hash files created with same settings perform very differently on the two DS versions. Whereas the 8.1 aix job completes in 15 mins the 8.7 version runs into hours.
Although the jobs sit in different test environments, both have very little running on them (and the jobs have been ran at various times with the same results) and I am led to believe should be similar in set up.
Jobs which read/write from flat file to flat file, to dev/null and to a 32 bit hash file (volumes just under two gig) on both 8.1 and 8.7 all perform at similar speeds.
The 64 bit hash file was created with the mkdbfile command.
The DS tunable settings appear to match.
As anyone experienced any such performance issues with 64 bit hash files on 8.7 or can anyone offer advice as to what to check that might specifically impact on jobs using 64 bit hash files?
Cheers.
Performance Test Problems with 64 bit Hash files on 8.7
Moderators: chulett, rschirm, roy
</a>I mean not after the mkdbfile but immediately after they were filled with data. If the source data is the same then both files should have exactly the same sizes immediately after filling.
If they are not identical that would point towards different configurations which, in turn, might explain the load time differences.
If the are identical then it might be due to physical things such as the target file system and underlying disk.
Rich - I'm not sailing yet, but am in the finishing phases of a project after which I'll relax in the sun on the boat again! How's Yorkshire?
If they are not identical that would point towards different configurations which, in turn, might explain the load time differences.
If the are identical then it might be due to physical things such as the target file system and underlying disk.
Rich - I'm not sailing yet, but am in the finishing phases of a project after which I'll relax in the sun on the boat again! How's Yorkshire?
Arnd,
I have cut the volumes down from 15 mill to 3 mill and the data and overflow files are exactly the same on both 8.7 and 8.1 test runs.
The job run time is twice as long (ran these jobs a few times) on 8.7. Its about 8,000 rps on 8.7 and 14,000 on 8.1. The startup time on the 8.1 job seems to be significamtly longer than on 8.7. Also the rps on the 8.1 job is fairly constant but on the 8.7 job it start out a bit lower than on the 8.1 job but then gradually falls away.
Any ideas?
I have cut the volumes down from 15 mill to 3 mill and the data and overflow files are exactly the same on both 8.7 and 8.1 test runs.
The job run time is twice as long (ran these jobs a few times) on 8.7. Its about 8,000 rps on 8.7 and 14,000 on 8.1. The startup time on the 8.1 job seems to be significamtly longer than on 8.7. Also the rps on the 8.1 job is fairly constant but on the 8.7 job it start out a bit lower than on the 8.1 job but then gradually falls away.
Any ideas?
No, I'm stymied as well for an explanation. You've stated that normal hashed file write speeds are similar and I'll assume that the files are on the same disk drive as the 64-bit ones so that should rule out different disk partition types/ I/O speeds being the culprit
Something is acting as a bottleneck here. Can you run the jobs and compare CPU loads on the two environments? If the job doesn't do much in the way of computation the CPU load shouldn't approach a high load but are the loads similar on both environments? Likewise can you monitor the I/O activity to the drive with the hashed files to see what the Mb/Sec loads are and if they are maxed out as well as seeing if the newer, slower 64Bit file is (for some reason) generating more I/O despite being identical in size to the original?
Something is acting as a bottleneck here. Can you run the jobs and compare CPU loads on the two environments? If the job doesn't do much in the way of computation the CPU load shouldn't approach a high load but are the loads similar on both environments? Likewise can you monitor the I/O activity to the drive with the hashed files to see what the Mb/Sec loads are and if they are maxed out as well as seeing if the newer, slower 64Bit file is (for some reason) generating more I/O despite being identical in size to the original?
have you verified the environment are equal ?
Did you check the UVCONFIG file for both ?
This reminds me of the UVSYNC and SYNCALOC variables. Changing from AIX to linux may have changed the default for these and thus dropping the performance when volume goes up.
Sorry for the delay, didn't see the post was old.
Maybe it helps anyway.
This reminds me of the UVSYNC and SYNCALOC variables. Changing from AIX to linux may have changed the default for these and thus dropping the performance when volume goes up.
Sorry for the delay, didn't see the post was old.
Maybe it helps anyway.