Re: [PATCH v2 0/8] vsock/virtio: optimizations to increase the throughput

From: Stefano Garzarella
Date: Mon May 13 2019 - 12:51:33 EST


On Mon, May 13, 2019 at 05:33:40PM +0800, Jason Wang wrote:
>
> On 2019/5/10 äå8:58, Stefano Garzarella wrote:
> > While I was testing this new series (v2) I discovered an huge use of memory
> > and a memory leak in the virtio-vsock driver in the guest when I sent
> > 1-byte packets to the guest.
> >
> > These issues are present since the introduction of the virtio-vsock
> > driver. I added the patches 1 and 2 to fix them in this series in order
> > to better track the performance trends.
> >
> > v1: https://patchwork.kernel.org/cover/10885431/
> >
> > v2:
> > - Add patch 1 to limit the memory usage
> > - Add patch 2 to avoid memory leak during the socket release
> > - Add patch 3 to fix locking of fwd_cnt and buf_alloc
> > - Patch 4: fix 'free_space' type (u32 instead of s64) [Stefan]
> > - Patch 5: Avoid integer underflow of iov_len [Stefan]
> > - Patch 5: Fix packet capture in order to see the exact packets that are
> > delivered. [Stefan]
> > - Add patch 8 to make the RX buffer size tunable [Stefan]
> >
> > Below are the benchmarks step by step. I used iperf3 [1] modified with VSOCK
> > support.
> > As Micheal suggested in the v1, I booted host and guest with 'nosmap', and I
> > added a column with virtio-net+vhost-net performance.
> >
> > A brief description of patches:
> > - Patches 1+2: limit the memory usage with an extra copy and avoid memory leak
> > - Patches 3+4: fix locking and reduce the number of credit update messages sent
> > to the transmitter
> > - Patches 5+6: allow the host to split packets on multiple buffers and use
> > VIRTIO_VSOCK_MAX_PKT_BUF_SIZE as the max packet size allowed
> > - Patches 7+8: increase RX buffer size to 64 KiB
> >
> > host -> guest [Gbps]
> > pkt_size before opt p 1+2 p 3+4 p 5+6 p 7+8 virtio-net + vhost
> > TCP_NODELAY
> > 64 0.068 0.063 0.130 0.131 0.128 0.188 0.187
> > 256 0.274 0.236 0.392 0.338 0.282 0.749 0.654
> > 512 0.531 0.457 0.862 0.725 0.602 1.419 1.414
> > 1K 0.954 0.827 1.591 1.598 1.548 2.599 2.640
> > 2K 1.783 1.543 3.731 3.637 3.469 4.530 4.754
> > 4K 3.332 3.436 7.164 7.124 6.494 7.738 7.696
> > 8K 5.792 5.530 11.653 11.787 11.444 12.307 11.850
> > 16K 8.405 8.462 16.372 16.855 17.562 16.936 16.954
> > 32K 14.208 13.669 18.945 20.009 23.128 21.980 23.015
> > 64K 21.082 18.893 20.266 20.903 30.622 27.290 27.383
> > 128K 20.696 20.148 20.112 21.746 32.152 30.446 30.990
> > 256K 20.801 20.589 20.725 22.685 34.721 33.151 32.745
> > 512K 21.220 20.465 20.432 22.106 34.496 36.847 31.096
> >
> > guest -> host [Gbps]
> > pkt_size before opt p 1+2 p 3+4 p 5+6 p 7+8 virtio-net + vhost
> > TCP_NODELAY
> > 64 0.089 0.091 0.120 0.115 0.117 0.274 0.272
> > 256 0.352 0.354 0.452 0.445 0.451 1.085 1.136
> > 512 0.705 0.704 0.893 0.858 0.898 2.131 1.882
> > 1K 1.394 1.433 1.721 1.669 1.691 3.984 3.576
> > 2K 2.818 2.874 3.316 3.249 3.303 6.719 6.359
> > 4K 5.293 5.397 6.129 5.933 6.082 10.105 9.860
> > 8K 8.890 9.151 10.990 10.545 10.519 15.239 14.868
> > 16K 11.444 11.018 12.074 15.255 15.577 20.551 20.848
> > 32K 11.229 10.875 10.857 24.401 25.227 26.294 26.380
> > 64K 10.832 10.545 10.816 39.487 39.616 34.996 32.041
> > 128K 10.435 10.241 10.500 39.813 40.012 38.379 35.055
> > 256K 10.263 9.866 9.845 34.971 35.143 36.559 37.232
> > 512K 10.224 10.060 10.092 35.469 34.627 34.963 33.401
> >
> > As Stefan suggested in the v1, this time I measured also the efficiency in this
> > way:
> > efficiency = Mbps / (%CPU_Host + %CPU_Guest)
> >
> > The '%CPU_Guest' is taken inside the VM. I know that it is not the best way,
> > but it's provided for free from iperf3 and could be an indication.
> >
> > host -> guest efficiency [Mbps / (%CPU_Host + %CPU_Guest)]
> > pkt_size before opt p 1+2 p 3+4 p 5+6 p 7+8 virtio-net + vhost
> > TCP_NODELAY
> > 64 0.94 0.59 3.96 4.06 4.09 2.82 2.11
> > 256 2.62 2.50 6.45 6.09 5.81 9.64 8.73
> > 512 5.16 4.87 13.16 12.39 11.67 17.83 17.76
> > 1K 9.16 8.85 24.98 24.97 25.01 32.57 32.04
> > 2K 17.41 17.03 49.09 48.59 49.22 55.31 57.14
> > 4K 32.99 33.62 90.80 90.98 91.72 91.79 91.40
> > 8K 58.51 59.98 153.53 170.83 167.31 137.51 132.85
> > 16K 89.32 95.29 216.98 264.18 260.95 176.05 176.05
> > 32K 152.94 167.10 285.75 387.02 360.81 215.49 226.30
> > 64K 250.38 307.20 317.65 489.53 472.70 238.97 244.27
> > 128K 327.99 335.24 335.76 523.71 486.41 253.29 260.86
> > 256K 327.06 334.24 338.64 533.76 509.85 267.78 266.22
> > 512K 337.36 330.61 334.95 512.90 496.35 280.42 241.43
> >
> > guest -> host efficiency [Mbps / (%CPU_Host + %CPU_Guest)]
> > pkt_size before opt p 1+2 p 3+4 p 5+6 p 7+8 virtio-net + vhost
> > TCP_NODELAY
> > 64 0.90 0.91 1.37 1.32 1.35 2.15 2.13
> > 256 3.59 3.55 5.23 5.19 5.29 8.50 8.89
> > 512 7.19 7.08 10.21 9.95 10.38 16.74 14.71
> > 1K 14.15 14.34 19.85 19.06 19.33 31.44 28.11
> > 2K 28.44 29.09 37.78 37.18 37.49 53.07 50.63
> > 4K 55.37 57.60 71.02 69.27 70.97 81.56 79.32
> > 8K 105.58 100.45 111.95 124.68 123.61 120.85 118.66
> > 16K 141.63 138.24 137.67 187.41 190.20 160.43 163.00
> > 32K 147.56 143.09 138.48 296.41 301.04 214.64 223.94
> > 64K 144.81 143.27 138.49 433.98 462.26 298.86 269.71
> > 128K 150.14 147.99 146.85 511.36 514.29 350.17 298.09
> > 256K 156.69 152.25 148.69 542.19 549.97 326.42 333.32
> > 512K 157.29 153.35 152.22 546.52 533.24 315.55 302.27
> >
> > [1] https://github.com/stefano-garzarella/iperf/
>
>
> Hi:
>
> Do you have any explanation that vsock is better here? Is this because of
> the mergeable buffer? If you, we need test with mrg_rxbuf=off.

Hi Jason,

virtio-net stays faster for packets with size up tp 16K/32K, maybe, as
you suggested, could be releated to mergeable buffer.

I'll try to disable it and re-run the tests.

Thanks,
Stefano