On Wed, Jul 04, 2018 at 01:29:50PM +0530, Kashyap Desai wrote: > Hi, > > Ming Lei posted below patch series and performance improved for > megaraid_sas driver. I used the same kernel base and figure out some more > possible performance improvement in block layer. This RFC improves > performance as well as CPU utilization. If this patch fits the design > aspect of the blk-mq and scsi-mq, I can convert it into PATCH and submit > the same/modified version. > > https://marc.info/?l=linux-block&m=153062994403732&w=2 > > Description of change - > > Do not insert request into software queue if BLK_MQ_F_NO_SCHED is set. > Submit request from blk_mq_make_request to low level driver directly as > depicted through below function call. > > blk_mq_try_issue_directly > > > __blk_mq_try_issue_directly > > scsi_queue_rq Hi Kashyap, When I sent you the patches[1] which include 'global tags' support, MegaRAID is converted to per-node queues and becomes real MQ(q->nr_hw_queues > 1), IO should have been issued direclty, please see the branch of '(plug && !blk_queue_nomerges(q))' and '(q->nr_hw_queues > 1 && is_sync)' in blk_mq_make_request(). But looks not see any improvement from your test result, even though huge improvement can be observed on null_blk/scsi_debug. Maybe somewhere is wrong, and days ago I talked with Laurence about doing this kind of test again. I will double check the 'global tags' patches, meantime could you or Laurence help to check if global tags[2] works in expected way if you'd like to? [1] https://github.com/ming1/linux/commits/v4.16-rc-host-tags-v5 [2] https://github.com/ming1/linux/commits/v4.18-rc-host-tags-v8 > > Low level driver attached to scsi.mq can set BLK_MQ_F_NO_SCHED, If they do > not want benefit from io scheduler (e.a in case of SSDs connected to IT/MR > controller). In case of HDD drives connected to HBA, driver can avoid > setting BLK_MQ_F_NO_SCHED so that default elevator is set to mq-deadline. That might be one way for improving megaraid_sas. However, even for modern high performance device, such as NVMe, turns out IO scheduler(such as kyber) may play a big role for improving latency or throughput. Thanks, Ming
