Messages in this thread | | | From | Rusty Russell <> | Subject | Re: [PATCH] blk-merge: fix blk_recount_segments | Date | Tue, 14 Oct 2014 14:24:10 +1030 |
| |
Ming Lei <ming.lei@canonical.com> writes: > Hi Rusty,
Hi Ming!
Sorry, I was on vacation. I'm back and slowly working through all my mail...
> 1, FIO script > [global] > direct=1 > size=128G > bsrange=${BS}-${BS} > timeout=60 > numjobs=4 > ioengine=libaio > iodepth=64 > filename=/dev/vdb #backed by /dev/nullb0, 4 virtqueues per virtio-blk > group_reporting=1 > > [f] > rw=randread
> 5, result > 5.1 without Rusty's virtio-vring patch > - BS=4K, throughput: 179K > - BS=256K, throughput: 27540
(ie. always using indirect)
> 5.2 with Rusty's virtio-vring patch > - BS=4K, throughput: 173K > - BS=256K, throughput: 25350
(ie. usually using indirect).
> Looks throughput decreases if BS is 256K in case of your patch.
Interesting. Looks like we're ending up with fewer descs in flight, though I'm surprised: with 256K blocks I'd expect us to hit the steady state of "all indirect" almost immediately.
Hmm... I think the heuristic in my patch is flawed. Let me try again, and get back to you.
Thanks, Rusty.
| |