From: Sebastian R. <seb...@pr...> - 2012-03-29 12:53:25
|
Hi list, we've got 3.2.8 kernels on both SRP target and initiator. The one on the target is patched with SCST 2.2.0. We use the ofa-kernel-modules from intree and the matching OFED-1.5.4 user space. Our QDR HCAs are the Mellanox MHQH19B-XTR (mlx4). On the storage there is a MD RAID-10 (MD RAID-1 + MD RAID-0) with 512 KiB stripe size over 16 HDDs. We do a file copy benchmark on ext4 without any caching which is pretty fast (big chunk sizes) local on the storage system. But over SRP it is much slower than with iSER and the tgt target. We're exporting simple partitions via SRP. On the initiator "iomin" is reported as 4 KiB and "ioopt" is reported as 512 KiB. blktrace shows that nothing bigger than 64 KiB is in the queue on the initiator as well as on the target with SRP. We've tried to increase that by setting the ib_srpt parameter "srp_max_rdma_size=524288", but it had no effect - still max. 64 KiB in the queue. We've also tried the performance tuning described in the README - didn't have much effect. How can we improve performance/increase chunk size there? Cheers, Sebastian -- Sebastian Riemer Linux Kernel Developer ProfitBricks GmbH Greifswalder Str. 207 10405 Berlin, Germany Tel.: +49 - 30 - 60 98 56 991 - 303 Fax: +49 - 30 - 51 64 09 22 Email: seb...@pr... Web: http://www.profitbricks.com/ Sitz der Gesellschaft: Berlin Registergericht: Amtsgericht Charlottenburg, HRB 125506 B Geschäftsführer: Andreas Gauger, Achim Weiss |