Description of problem: I am grappling with IO bottleneck issues on Xeon servers running RHEL3. They seem to max out IO capacity at 30-40 MB/s. Some say that's expected, but no normal ranges are anywere to be found. Still I think there is room for improvement. We do 2 Gb/s FibreChannel to an EMC Symmetrix and I have the EMC engineer make statistics of the EMC device while I do a simple read test with dd bs=65536 if=/dev/sda1 of=/dev/null I also alter vm.max-readahead in two test to be 1023 and 127 respectively. I expect to see the read sizes on the EMC to change. - I expect them to be 64 kB minimum, because I set read buffer size to 65536 on dd. - I expect them to be 64 kB when vm.max-readahead is 127. - I expect them to grow up to 512 kB when I set vm.max- readahead to 1023 Yet, in all cases the EMC gets read requests of 32 kB only. Question: - is there a hard limit in the linux kernel for IO buffers? - is there a hard limit in the QLogic 2200 driver? - is there a limit on the QLogic card itself? - is there a sysctl parameter that I can set to crank the maximum transfer size up to at least 128 kB? Version-Release number of selected component (if applicable): kernel: 2.4.21-15.0.3.ELsmp QLA driver version 6.07.02-RH2 How reproducible: Steps to Reproduce: see above Actual results: 32kB transfer sizes on the SAN disk array. Expected results: 64kB and 512kB transfer sizes on the SAN disk array.
I need to add that there is no hard limit configured on the HBA, I just checked that. So, the HBA is not doing the fragmenting. The EMC people say they like to see 128kB request sizes, so it's not them breaking it up into 32k fragments either.
Will there be any action on this? I have submitted this weeks ago and no news at all? Thanks for your attention.
This problem was likely from a scatter gather limit in the QLogic driver. Given the QLogic driver has been updated several times since this bug report, does this problem still exist with the latest RHEL3 update? If not, it should be closed.
Given where RHEL 3 is in it's lifecycle, the age of the last customer comment, plus the possibility the QLA driver might have solved this issue, I'm closing this out. Re-open if you feel differently.