Re: [PATCH 2/2] scsi-disk: Add native FUA support

[Kevin Wolf]

Am 06.03.2025 um 11:33 hat Kevin Wolf geschrieben:
> Am 04.03.2025 um 16:52 hat Alberto Faria geschrieben:
> > Avoid emulating FUA when the driver supports it natively. This should
> > provide better performance than a full flush after the write.
> > 
> > Signed-off-by: Alberto Faria <afaria@redhat.com>
> 
> Did you try out if you can see performance improvements in practice?
> It's always nice to have numbers in the commit message for patches that
> promise performance improvements.

I was curious enough to see how this and the recent series by Stefan
(virtio-scsi multiqueue) and myself (FUA on the backend + polling
improvements) play out with virtio-scsi, so I just ran some fio
benchmarks with sync=1 myself to compare:

iops bs=4k cache=none           |    virtio-scsi    |     virtio-blk    |
O_SYNC workload                 |   qd 1  |  qd 16  |   qd 1  |  qd 16  |
--------------------------------+---------+---------+---------+---------+
master                          |   21296 |  109747 |   25762 |  130576 |
+ virtio-scsi multiqueue        |   28798 |  121170 |       - |       - |
+ FUA in scsi-disk              |   51893 |  204199 |       - |       - |
--------------------------------+---------+---------+---------+---------+
Total change                    | +143.7% |  +86.1% |       - |       - |

(No new numbers for virtio-blk because virtio-scsi patches obviously
don't change anything about it. Also no numbers for FUA in file-posix
because it's unused with cache=none.)

iops bs=4k cache=directsync     |    virtio-scsi    |     virtio-blk    |
O_SYNC workload                 |   qd 1  |  qd 16  |   qd 1  |  qd 16  |
--------------------------------+---------+---------+---------+---------+
master                          |   32223 |  109748 |   45583 |  258416 |
+ FUA in file-posix + polling   |   32148 |  198665 |   58601 |  320190 |
+ virtio-scsi multiqueue        |   51739 |  225031 |       - |       - |
+ FUA in scsi-disk              |   56061 |  227535 |       - |       - |
--------------------------------+---------+---------+---------+---------+
Total change                    |  +74.0% | +107.3% |  +28.6% |  +23.9% |

Of course, the huge improvements on the virtio-scsi side only show how
bad it was before. In most numbers it is still behind virtio-blk even
after all three patch series (apart from cache=none where the
availability of FUA on the device side makes a big difference, and I
expect that virtio-blk will improve similarly once we implement it
there).

Also note that when testing the virtio-scsi multiqueue patches, this
was still a single iothread, i.e. I wasn't even making use of the new
feature per se. I assume much of this comes from enabling polling
because the series moved the event queue handling to the main loop,
which prevented polling for virtio-scsi before. The series also got rid
of an extra coroutine per request for the blk_is_available() call in
virtio_scsi_ctx_check(), which might play a role, too.

Anyway, I like these numbers for FUA in scsi-disk. It makes write back
cache modes almost catch up to write through with O_SYNC workloads. We
should definitely get this merged and do the same for virtio-blk.

Kevin