On Wed, Jul 08, 2020 at 06:59:41AM -0400, Michael S. Tsirkin wrote:
> On Mon, Jul 06, 2020 at 02:56:45PM +0100, Stefan Hajnoczi wrote:
> > v4:
> > * Sorry for the long delay. I considered replacing this series with a simpler
> > approach. Real hardware ships with a fixed number of queues (e.g. 128). The
> > equivalent can be done in QEMU too. That way we don't need to magically size
> > num_queues. In the end I decided against this approach because the Linux
> > virtio_blk.ko and virtio_scsi.ko guest drivers unconditionally initialized
> > all available queues until recently (it was written with
> > num_queues=num_vcpus in mind). It doesn't make sense for a 1 CPU guest to
> > bring up 128 virtqueues (waste of resources and possibly weird performance
> > effects with blk-mq).
> > * Honor maximum number of MSI-X vectors and virtqueues [Daniel Berrange]
> > * Update commit descriptions to mention maximum MSI-X vector and virtqueue
> > caps [Raphael]
> > v3:
> > * Introduce virtio_pci_optimal_num_queues() helper to enforce VIRTIO_QUEUE_MAX
> > in one place
> > * Use VIRTIO_SCSI_VQ_NUM_FIXED constant in all cases [Cornelia]
> > * Update hw/core/machine.c compat properties for QEMU 5.0 [Michael]
> > v3:
> > * Add new performance results that demonstrate the scalability
> > * Mention that this is PCI-specific [Cornelia]
> > v2:
> > * Let the virtio-DEVICE-pci device select num-queues because the optimal
> > multi-queue configuration may differ between virtio-pci, virtio-mmio, and
> > virtio-ccw [Cornelia]
> >
> > Enabling multi-queue on virtio-pci storage devices improves performance on SMP
> > guests because the completion interrupt is handled on the vCPU that submitted
> > the I/O request. This avoids IPIs inside the guest.
> >
> > Note that performance is unchanged in these cases:
> > 1. Uniprocessor guests. They don't have IPIs.
> > 2. Application threads might be scheduled on the sole vCPU that handles
> > completion interrupts purely by chance. (This is one reason why benchmark
> > results can vary noticably between runs.)
> > 3. Users may bind the application to the vCPU that handles completion
> > interrupts.
> >
> > Set the number of queues to the number of vCPUs by default on virtio-blk and
> > virtio-scsi PCI devices. Older machine types continue to default to 1 queue
> > for live migration compatibility.
> >
> > Random read performance:
> > IOPS
> > q=1 78k
> > q=32 104k +33%
> >
> > Boot time:
> > Duration
> > q=1 51s
> > q=32 1m41s +98%
> >
> > Guest configuration: 32 vCPUs, 101 virtio-blk-pci disks
> >
> > Previously measured results on a 4 vCPU guest were also positive but showed a
> > smaller 1-4% performance improvement. They are no longer valid because
> > significant event loop optimizations have been merged.
>
> I'm guessing this should be deferred to the next release as
> it (narrowly) missed the freeze window. Does this make sense to you?
Yes, that is fine. Thanks!
Stefan