On Thu, Mar 08, 2018 at 08:52:52AM +0100, Christoph Hellwig wrote: > On Tue, Feb 27, 2018 at 06:07:46PM +0800, Ming Lei wrote: > > This patch can support to partition host-wide tags to multiple hw queues, > > so each hw queue related data structures(tags, hctx) can be accessed in > > NUMA locality way, for example, the hw queue can be per NUMA node. > > > > It is observed IOPS can be improved much in this way on null_blk test. > > null_blk isn't too interesting, so some real hardware number would > be very useful here.
About 10~20% IOPS improvement can be observed on scsi_debug too, which is setup on one dual-sockets system. It needs one hpsa or megaraid_sas host with dozens of SSDs, which seems not easy to setup for me. And Kashyap is very cooperative to test patches, looks V3 is much better than before by using per-node hw queue. If atomic operations on scsi_host->host_busy are removed, and megaraid_sas IO path can be optimized a bit, we should get some improvement by per-node hw queue with BLK_MQ_F_HOST_TAGS on megaraid_sas. > > Also the documentation should be a lot less sparse. When are we going > to set this flag? What help are we going to give driver authors to > guide chosing the option? OK, will do that in next version. Thanks, Ming
