Re: [Qemu-devel] Measure virtio-vsock speed

[Stefan Hajnoczi]

On Thu, Dec 15, 2016 at 05:00:35PM +0800, Zhang Shuaiyi wrote:
> I used kernel 4.8 and qemu v2.8.0-rc3.
> Qemu command line:
>   -device vhost-vsock-pci,id=vhost-vsock-pci0,guest-cid=3
> 
> I try to use nc-vsock to test vsock speed.
> 1. In host:
>   # modprobe vhost-vsock
>   # ./nc-vsock -l 1234 > /dev/null
> 2. In guest:
>   a. use pv and nc-vsock, speed is about 62MB/s.
>      # pv /dev/zero | ./nc-vsock 3 1234
>   b. use dd and nc-vsock, speed is about 67MB/s.
>      # dd if=/dev/zero bs=1M count=2000 | ./nc-vsock 3 1234
>      
> I think nc-vsock can not get the maximum speed, so I modified
> nc-vsock.c.

Good idea, nc-vsock is not a performance benchmark.

> 1. In host, just recv data without printf.
> 2. In guest, use gettimeofday to record interval and send zero bufs.
>   #define MAXLEN 65536
>   char buf[MAXLEN] = {0};
>   unsigned long long num = 0;
> 
>   gettimeofday(&start_time, NULL);
>   for (n = 0; n  1000000; n++) {
>       ret = send(fd, buf, MAXLEN, 0);
>       if (ret  0) {
>          fprintf(stderr, "send error ret:%d\n", ret);
>          break;
>       }
>       num += ret;
>   }
>   gettimeofday(&end_time, NULL);

This is benchmarking how quickly send(2) returns but it doesn't
guarantee that the guest has received everything.  Is it necessary to
add a final recv(fd, &status, sizeof(status), 0) to check that the other
side has seen all the data?

> In this way, the virtio-vsock speed is about 500MB/s. But use iperf to
> test virtio-net, the speed is about 1.5GB/s.

Your vsock benchmark may not be doing the same thing as iperf.  I
suggest adding AF_VSOCK support to iperf for a fair comparison.

> Iperf can be used to test virtio-net, is there a tool to test maximum 
> virtio-vsock speed?
> At present, virtio-vsock is faster than virtio-net?

I expect virtio-net to be faster than virtio-vsock and to remain faster
in the future.  The goal of vsock is to provide reliable,
zero-configuration communication between guest and host.  The goal is
not to be a faster transport than virtio-net.

If you are considering running something over AF_VSOCK for performance
reasons then that's not what it was designed for.  If performance is the
*most* important thing look at vhost-user (i.e. shared memory) but it
has a lot of trade-offs because you need to implement virtio-net and a
network stack.

signature.asc
Description: PGP signature