[Qemu-devel] Re: [PATCH] virtio: Use ioeventfd for virtqueue notify

[Avi Kivity <avi@redhat.com>]

  On 09/30/2010 04:01 PM, Stefan Hajnoczi wrote:
> Virtqueue notify is currently handled synchronously in userspace virtio.
> This prevents the vcpu from executing guest code while hardware
> emulation code handles the notify.
>
> On systems that support KVM, the ioeventfd mechanism can be used to make
> virtqueue notify a lightweight exit by deferring hardware emulation to
> the iothread and allowing the VM to continue execution.  This model is
> similar to how vhost receives virtqueue notifies.

Note that this is a tradeoff.  If an idle core is available and the 
scheduler places the iothread on that core, then the heavyweight exit is 
replaced by a lightweight exit + IPI.  If the iothread is co-located 
with the vcpu, then we'll take a heavyweight exit in any case.

The first case is very likely if the host cpu is undercommitted and 
there is heavy I/O activity.  This is a typical subsystem benchmark 
scenario (as opposed to a system benchmark like specvirt).  My feeling 
is that total system throughput will be decreased unless the scheduler 
is clever enough to place the iothread and vcpu on the same host cpu 
when the system is overcommitted.

We can't balance "feeling" against numbers, especially when we have a 
precedent in vhost-net, so I think this should go in.  But I think we 
should also try to understand the effects of the extra IPIs and 
cacheline bouncing that this creates.  While virtio was designed to 
minimize this, we know it has severe problems in this area.

> The result of this change is improved performance for userspace virtio
> devices.  Virtio-blk throughput increases especially for multithreaded
> scenarios and virtio-net transmit throughput increases substantially.
> Full numbers are below.
>
> This patch employs ioeventfd virtqueue notify for all virtio devices.
> Linux kernels pre-2.6.34 only allow for 6 ioeventfds per VM and care
> must be taken so that vhost-net, the other ioeventfd user in QEMU, is
> able to function.  On such kernels ioeventfd virtqueue notify will not
> be used.
>
> Khoa Huynh<khoa@us.ibm.com>  collected the following data for
> virtio-blk with cache=none,aio=native:
>
> FFSB Test          Threads  Unmodified  Patched
>                              (MB/s)      (MB/s)
> Large file create  1        21.7        21.8
>                     8        101.0       118.0
>                     16       119.0       157.0
>
> Sequential reads   1        21.9        23.2
>                     8        114.0       139.0
>                     16       143.0       178.0
>
> Random reads       1        3.3         3.6
>                     8        23.0        25.4
>                     16       43.3        47.8
>
> Random writes      1        22.2        23.0
>                     8        93.1        111.6
>                     16       110.5       132.0

Impressive numbers.  Can you also provide efficiency (bytes per host cpu 
seconds)?

How many guest vcpus were used with this?  With enough vcpus, there is 
also a reduction in cacheline bouncing, since the virtio state in the 
host gets to stay on one cpu (especially with aio=native).

> Sridhar Samudrala<sri@us.ibm.com>  collected the following data for
> virtio-net with 2.6.36-rc1 on the host and 2.6.34 on the guest.
>
> Guest to Host TCP_STREAM throughput(Mb/sec)
> -------------------------------------------
> Msg Size  vhost-net  virtio-net  virtio-net/ioeventfd
> 65536         12755        6430                  7590
> 16384          8499        3084                  5764
>   4096          4723        1578                  3659
>   1024          1827         981                  2060

Even more impressive (expected since the copying, which isn't present 
for block, is now shunted off into an iothread).

On the last test you even exceeded vhost-net.  Any theories how/why?

Again, efficiency numbers would be interesting.

> Host to Guest TCP_STREAM throughput(Mb/sec)
> -------------------------------------------
> Msg Size  vhost-net  virtio-net  virtio-net/ioeventfd
> 65536         11156        5790                  5853
> 16384         10787        5575                  5691
>   4096         10452        5556                  4277
>   1024          4437        3671                  5277

Here you exceed vhost-net, too.

> +static int kvm_check_many_iobus_devs(void)
> +{
> +    /* Older kernels have a 6 device limit on the KVM io bus.  In that case
> +     * creating many ioeventfds must be avoided.  This tests checks for the
> +     * limitation.
> +     */
> +    EventNotifier notifiers[7];
> +    int i, ret = 0;
> +    for (i = 0; i<  ARRAY_SIZE(notifiers); i++) {
> +        ret = event_notifier_init(&notifiers[i], 0);
> +        if (ret<  0) {
> +            break;
> +        }
> +        ret = kvm_set_ioeventfd_pio_word(event_notifier_get_fd(&notifiers[i]), 0, i, true);
> +        if (ret<  0) {
> +            event_notifier_cleanup(&notifiers[i]);
> +            break;
> +        }
> +    }
> +
> +    /* Decide whether many devices are supported or not */
> +    ret = i == ARRAY_SIZE(notifiers);
> +
> +    while (i-->  0) {
> +        kvm_set_ioeventfd_pio_word(event_notifier_get_fd(&notifiers[i]), 0, i, false);
> +        event_notifier_cleanup(&notifiers[i]);
> +    }
> +    return ret;
> +}

Sorry about that.

IIRC there was a problem (shared by vhost-net) with interrupts remaining 
enabled in the window between the guest kicking the queue and the host 
waking up and disabling interrupts.  An even more vague IIRC mst had an 
idea to fix this?

-- 
error compiling committee.c: too many arguments to function