Re: [Qemu-devel] [PATCH v4 12/13] aio: self-tune polling time

[Stefan Hajnoczi]

On Mon, Dec 05, 2016 at 09:06:17PM +0100, Christian Borntraeger wrote:
> On 12/01/2016 08:26 PM, Stefan Hajnoczi wrote:
> > This patch is based on the algorithm for the kvm.ko halt_poll_ns
> > parameter in Linux.  The initial polling time is zero.
> > 
> > If the event loop is woken up within the maximum polling time it means
> > polling could be effective, so grow polling time.
> > 
> > If the event loop is woken up beyond the maximum polling time it means
> > polling is not effective, so shrink polling time.
> > 
> > If the event loop makes progress within the current polling time then
> > the sweet spot has been reached.
> > 
> > This algorithm adjusts the polling time so it can adapt to variations in
> > workloads.  The goal is to reach the sweet spot while also recognizing
> > when polling would hurt more than help.
> > 
> > Two new trace events, poll_grow and poll_shrink, are added for observing
> > polling time adjustment.
> > 
> > Signed-off-by: Stefan Hajnoczi <address@hidden>
> 
> Not sure way, but I have 4 host ramdisks with the same iothread as guest
> virtio-blk. running fio in the guest on one of these disks will poll, as
> soon as I have 2 disks  in fio I almost always see shrinks (so polling 
> stays at 0) and almost no grows.

Shrinking occurs when polling + ppoll(2) time exceeds poll-max-ns.

What is the value of poll-max-ns and how long is run_poll_handlers_end -
run_poll_handlers_begin?

I wonder if polling both disks takes longer than poll-max-ns once you
have two disks.  The "polling" activity includes processing the I/O
requests, so I imagine the time extends significantly as more disks have
I/O requests ready for processing.

Maybe the block_ns timing calculation should exclude processing time to
avoid false shrinking?

It also strikes me that there's a blind spot to the self-tuning
algorithm: imagine virtqueue kick via ppoll(2) + ioeventfd takes N
nanoseconds.  Detecting new virtqueue buffers via polling takes M
nanoseconds.  When M <= poll-max-ns < N the algorithm decides there is
no point in polling but it would actually be faster to poll.  The reason
is that the algorithm only looks at block_ns, which is N, not M.

This seems difficult to tackle because the algorithm has no way of
predicting M unless it randomly tries to poll longer.

Stefan

signature.asc
Description: PGP signature