Re: [RFC] thread-pool: Add option to fix the pool size

[Stefan Hajnoczi]

On Thu, Feb 03, 2022 at 10:56:49AM +0000, Daniel P. Berrangé wrote:
> On Thu, Feb 03, 2022 at 10:53:07AM +0000, Stefan Hajnoczi wrote:
> > On Wed, Feb 02, 2022 at 06:52:34PM +0100, Nicolas Saenz Julienne wrote:
> > > The thread pool regulates itself: when idle, it kills threads until
> > > empty, when in demand, it creates new threads until full. This behaviour
> > > doesn't play well with latency sensitive workloads where the price of
> > > creating a new thread is too high. For example, when paired with qemu's
> > > '-mlock', or using safety features like SafeStack, creating a new thread
> > > has been measured take multiple milliseconds.
> > > 
> > > In order to mitigate this let's introduce a new option to set a fixed
> > > pool size. The threads will be created during the pool's initialization,
> > > remain available during its lifetime regardless of demand, and destroyed
> > > upon freeing it. A properly characterized workload will then be able to
> > > configure the pool to avoid any latency spike.
> > > 
> > > Signed-off-by: Nicolas Saenz Julienne <nsaenzju@redhat.com>
> > > 
> > > ---
> > > 
> > > The fix I propose here works for my specific use-case, but I'm pretty
> > > sure it'll need to be a bit more versatile to accommodate other
> > > use-cases.
> > > 
> > > Some questions:
> > > 
> > > - Is unanimously setting these parameters for any pool instance too
> > >   limiting? It'd make sense to move the options into the AioContext the
> > >   pool belongs to. IIUC, for the general block use-case, this would be
> > >   'qemu_aio_context' as initialized in qemu_init_main_loop().
> > 
> > Yes, qemu_aio_context is the main loop's AioContext. It's used unless
> > IOThreads are configured.
> > 
> > It's nice to have global settings that affect all AioContexts, so I
> > think this patch is fine for now.
> > 
> > In the future IOThread-specific parameters could be added if individual
> > IOThread AioContexts need tuning (similar to how poll-max-ns works
> > today).
> > 
> > > - Currently I'm setting two pool properties through a single qemu
> > >   option. The pool's size and dynamic behaviour, or lack thereof. I
> > >   think it'd be better to split them into separate options. I thought of
> > >   different ways of expressing this (min/max-size where static happens
> > >   when min-size=max-size, size and static/dynamic, etc..), but you might
> > >   have ideas on what could be useful to other use-cases.
> > 
> > Yes, "min" and "max" is more flexible than fixed-size=n. fixed-size=n is
> > equivalent to min=n,max=n. The current default policy is min=0,max=64.
> > If you want more threads you could do min=0,max=128. If you want to
> > reserve 1 thread all the time use min=1,max=64.
> > 
> > I would go with min and max.
> 
> This commit also exposes this as a new top level command line
> argument. Given our aim to eliminate QemuOpts and use QAPI/QOM
> properties for everything I think we need a different approach.
> 
> I'm not sure which exisiting QAPI/QOM option it most appropriate
> to graft these tunables onto ?  -machine ?  -accel ?  Or is there
> no good fit yet ?

Yep, I didn't comment on this because I don't have a good suggestion.

In terms of semantics I think we should have:

1. A global default value that all new AioContext take. The QEMU main
   loop's qemu_aio_context will use this and all IOThread AioContext
   will use it (unless they have been overridden).

   I would define it on --machine because that's the "global" object for
   a guest, but that's not very satisfying.

2. (Future patch) --object iothread,thread-pool-min=N,thread-pool-max=M
   just like poll-max-ns and friends. This allows the values to be set
   on a per-IOThread basis.

Stefan

signature.asc
Description: PGP signature