Re: [Qemu-devel] [PATCH v2 2/2] block: bump coroutine pool size for drives

[Markus Armbruster]

Stefan Hajnoczi <address@hidden> writes:

> On Fri, Jul 04, 2014 at 12:03:27PM +0200, Markus Armbruster wrote:
>> Stefan Hajnoczi <address@hidden> writes:
>> > @@ -2112,6 +2115,7 @@ void bdrv_detach_dev(BlockDriverState *bs, void *dev)
>> >      bs->dev_ops = NULL;
>> >      bs->dev_opaque = NULL;
>> >      bs->guest_block_size = 512;
>> > +    qemu_coroutine_adjust_pool_size(-64);
>> >  }
>> >  
>> >  /* TODO change to return DeviceState * when all users are qdevified */
>> 
>> This enlarges the pool regardless of how the device model uses the block
>> layer.  Isn't this a bit crude?
>> 
>> Have you considered adapting the number of coroutines to actual demand?
>> Within reasonable limits, of course.
>
> I picked the simplest algorithm because I couldn't think of one which is
> clearly better.  We cannot predict future coroutine usage so any
> algorithm will have pathological cases.

Dynamically adapting to actual usage arguably involves less predicting
than 64 * #block backends.  Grow the pool some when we're running out of
coroutines, shrink it some when it's been underutilized for some time.

> In this case we might as well stick to the simplest implementation.

Keeping it simple is always a weighty argument.