[Ksummit-2012-discuss] [ATTEND] target + scsi mixed mode interaction, scsi IOPs performance limitations, tcm_vhost shared storage ALUA multipath

Nicholas A. Bellinger nab at linux-iscsi.org
Mon Jun 18 00:56:50 UTC 2012


Hello KS-2012 PCs,

I would like to attend KS-2012 this year in San Diego, and think I can
benefit the discussion in the following ways..

I'm maintainer of the target storage subsystem merged in .38 code using
the first fabric independent inter-module configfs control plane.  With
lots of help of target developers (hch, agrover, hannes), other
subsystem maintainers (roland, greg-kh, james, felipe, stefan), and new
devels (bart, bootc, sebastian) the team has manage to merge seven
fabric drivers, including three new drivers this round for v3.5 code
(usb-gadget, sbp-target, tcm_qla2xxx). 

For the 3.6 -> 3.7 time frame, one of the bigger items still on the TODO
list is how collectively the scsi <-> target subsystems are going to be
able to support "hot" target/initiator mixed-mode operation.  There has
been some discussion on this topic on linux-scsi + target-devel, but
we've not decided on a path forward here yet.. 

Also, another important topic for scsi that will need to eventually be
addressed is the per-lun performance limitations / bottleneck currently
present with v3.x Linux/SCSI clients for random small block IOPs
workloads.  Even after SCSI tree wide host-lock less conversion starting
in .38 code, Linux/SCSI client LUNs (regardless of fabric / LLD) are
still well behind what a raw block client is capable for 1 million IOPs
machine.  The Linux/SCSI client is also behind (on the order of ~25%) of
what certain non Linux client can achieve connected to the same Linux
target using raw block backends in an apples-to-apples comparison.

Finally on the new target fabric driver front, we're currently expecting
the tcm_vhost driver to be a merge candidate in the v3.6 -> v3.6 time
frame. This driver provides high performance backend (pSCSI, IBLOCK,
FILEIO) device access using existing vhost infrastructure for connecting
para-virtualized virtio-scsi guest code directly to the in-kernel target
infrastructure.

This is some really pretty exciting code, and I'll be sending time on
virto-scsi <-> tcm_vhost performance profiling using raw block flash
backends to get a rough idea of where things currently stand..  Also by
the time KS/LC comes around in August, we expect to have the first
shared-storage host virtualized ALUA multipath prototype up and running
on mainline target code..

Thank you!

--nab



More information about the Ksummit-2012-discuss mailing list