From: Allen, W. <way...@in...> - 2010-08-16 15:27:50
|
Hi Neto, Since the original IOMeter Linux implementation there have been some Linux kernel changes that have broken queued IO. You should be able to repro non-queued performance fairly close between Windows and Linux (they will never match exactly though). We have this on our list of feature enhancements to make IOMeter compatible with the current Linux generation. If you urgently need a solution for Linux, FIO is a tool that you could use. Best Regards, Wayne From: is...@ya... [mailto:is...@ya...] Sent: Sunday, August 15, 2010 6:49 PM To: iom...@li...; iom...@li... Subject: [Iometer-devel] Linux (dynamo) strange behavior: Strange performance numbers and bad performance using reads Hi All, This is neto from Brazil How are you? First of All, I would like to thank you Joe, Ved and Daniel for all support and help. Now, my question is regarding: Linux (dynamo) behavior: Strange performance numbers and bad performance using reads I'm using latest Branch (but I already have had the same behavior with other Iometer versions) Windows - GUI Linux - dynamo (Redhat 5.4 using dm-mp (multipathing) - Server with 64GB RAM. Using FC (Fibre Channel) to connect to Storage Array. Using writes (random writes) - performance is OK but the I/O numbers reported by Iometer are not accurate. (Storage reports 500 IOPS and Iometer reports 200 IOPS) Using reads (random reads) - performance is terrible. Using one worker or 50 workers the performance is the same. Increasing outstanding I/O doesn't help. I have tried to recompile using O_DIRECT - same behavior. Questions: 1) Why the numbers are not accurate? 2) Why do I have this "poor" performance on Linux (reads)? Writes are OK. 3) Outstanding I/O doesn't work on Linux. Why? IMPORTANT: On Windows (dynamo) everything works as expected. :-) Could you please help me to address those "questions"? Thank you very much All the best neto |