-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA256 On 03/31/2015, 03:30 PM, Jiri Slaby wrote:
On 03/24/2015, 03:45 PM, Jeff Mahoney wrote:
On 3/24/15 9:20 AM, David Sterba wrote:
Hi,
currently the scsi-mq is set to Y, since 3.18-rc1 merge. The option enables a feature that makes advantage of enterprise-class storage. It's known not to perform well on slower devices [1] and lacks scheduler support.
That's overstating the problem a bit. It lacks scheduler support because as we encounter devices capable of much higher numbers of IOPS with essentially zero seek latency, the time we spend processing requests starts to take longer than the I/O itself. This isn't limited to enterprise storage. I can order a commodity SSD this morning that meets this description for about the same price as a regular hard drive (obviously with less capacity.)
We'll set the option to N in our stable kernels.
The lack of scheduler support can cause significant performance drop. Vojtech reports 10-50x slowdown on random read workloads, the disks seek all over the platters due to the missing io scheduler optimizations.
As most of our users are not likely to run on high-end storage devices or otherwise benefit from the scsi-mq feature, I believe it's safe to turn the scsi-mq option off. Once the io scheduler support is availabe we can enable it again.
The change will be done for all architectures and all kernel flavors (30 changes in total).
I disagree. This is a scenario that benefits some hardware and costs other hardware, which is effectively the same scenario that exists without SCSI_MQ_DEFAULT. The difference is that the single queue block layer is dying and the multiqueue block layer is evolving.
The config option is the default and it can be changed at runtime. We're talking about Factory. If users with Vojtech's configuration (which is pretty esoteric - it's more than just having a few spinning disks) need to back off scsi-mq, they can do it by booting with scsi_mod.use_blk_mq=0. There is basic infrastructure in place to choose old block or blk-mq on a per-host basis but it hasn't been fleshed out yet.
David?
Ping: this is a month now. I will revert the change in master and merge to stable if there are no objections?
It has been a week now and it remained disabled. So what is the result of this thread, please?
thanks,
- -- js suse labs -----BEGIN PGP SIGNATURE----- Version: GnuPG v2 iQIcBAEBCAAGBQJVOJorAAoJEL0lsQQGtHBJ+ZsP/0pB115iSrKxSk12Lm6M1O+4 ybTdmHqhBcR4sd3RRhzL8+CcbV5E2v/OB6gn2K06YUYZW83dbsHeE8Rd53xlbtj5 NjsDbYBrMjBkDSVb3QgkyIPaOs6S6ToaVsz8ffV9EQuKNTbHyg72k3cOHpxxMX8N iyS9ozGivTPnS+G97Vj/J37/mUtC0Jv5vb7gUioo8A2/DhxxsTpWzaqiVSFgFAsP wvlMXgR0UeoWEQ3m65zNI3Jvh3zQ1m5zEI2uPCu8bp+WEgml37WgZs3vszKduUV/ bwy/LwqIKacMV2cXE0P3JAA9fM0ErQB9Y1n2kpVC+vnN+f0JfnaEufEpNBR16NH3 qr73MY47XB49oDIv1A6lBwps0N3CysmyMeZ5bNqwS6ciqpQuszPMNnGA4Mhl+3me pz8NWbUlknt8wnfuWRgPpNDgVg+qQLJkwIYeNRNBHgLjM3MmLfFX2Rkg/f8sa6Wf HcNqrKxFjk80plRLQzs/JZEz/3Qv7QMvIR2Gxloy2Pp7qvGAr+QqIWD5bmsig2a8 a8zsBdgvu7w4JVRnms+FgDICYsisSMh4/Xcu5v2f6/RCktgTzIjRXHNHNdZHKejB rH8tsm4SOgDRFWiL4eeCsA3AaNqE3mT4EDjNCBnFrG21WjHbhzpxkGl9fyKY8cl7 j/QHuzks4vgkSzP/XIg4 =229f -----END PGP SIGNATURE----- -- To unsubscribe, e-mail: opensuse-kernel+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse-kernel+owner@opensuse.org