Sorry, one last comment on issue #1 (slow with SCST iSCSI but fast qla2xxx FC with Ceph RBD):
tly work fine in combination with SCST so I'd recommend to continue testing with a recent kernel. I'm running myself kernel 4.3.0 since some time on my laptop and development workstation.I have succeeded in compiling SCST latest code with gcc 4.9 on Ubuntu - thank you. However, now I am seeing some strange behaviors:1. Same block device presented via blockio handler to ESXi performs very differently over fibre channel (QLE2562 HBA with latest firmware and qla2xxx SCST module) vs. iSCSI - fibre channel is much faster and on a time graph iSCSI is much choppier. 60000 IOPS on FC and 10000 IOPS on iSCSI using the same back end device (ceph rbd)
Any chance this is related to the blk-mq related concerns that Ilya Dryomov addressed in this discussion? http://www.spinics.net/lists/ceph-users/msg24394.html
Performance of the iSCSI device using the RBD backend is very similar to performance of the RBD backend alone with 4.1 kernels. With 4.2 kernels the RBD backend on its own performs great (dd tests show that very well), so perhaps the FC module uses the write combining mechanism and iSCSI does not?
Thank you for any insight - cross posting to ceph-users for any comment.
Best regards,
ALex
_______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
