On latest kernels, for most HDDs (for me actual spinning with SCSI interface -
IDE/SATA in AHCI mode & Megaraid) IO schedulers can be replaced by blk-mq
per-CPU queue. Even I put one node with 3.18 kernel into this mode (Megaraid,
scsi_mod.use_blk_mq=Y), planned to switch all nodes (include AHCI).
So, I have ceph-specific question. Are you prognosed this way effective,
according to multiple op threads (I have default 2)? Or in other words: will
ceph too frequental split one sequental client stream into multiple op threads
inside one PG (and significant reduce merging per-CPU, cause perfomance
degradation) or there will be mostly one thread?
--
WBR, Dzianis Kahanovich AKA Denis Kaganovich, http://mahatma.bspu.unibel.by/
_______________________________________________
ceph-users mailing list
[email protected]
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
