On Fri, Jun 29, 2012 at 11:08:18AM +0200, Michal Privoznik wrote:

On 29.06.2012 08:09, Eiichi Tsukata wrote:

...

Hello,

This is a patch to fix vm's outbound traffic control problem.

Currently, vm's outbound traffic control by libvirt doesn't go well. This problem was previously discussed at libvir-list ML, however it seems that there isn't still any answer to the problem. http://www.redhat.com/archives/libvir-list/2011-August/msg00333.html

I measured Guest(with virtio-net) to Host TCP throughput with the command "netperf -H". Here are the outbound QoS parameters and the results.

outbound average rate[kilobytes/s] : Guest to Host throughput[Mbit/s] ====================================================================== 1024 (8Mbit/s) : 4.56 2048 (16Mbit/s) : 3.29 4096 (32Mbit/s) : 3.35 8192 (64Mbit/s) : 3.95 16384 (128Mbit/s) : 4.08 32768 (256Mbit/s) : 3.94 65536 (512Mbit/s) : 3.23

The outbound traffic goes down unreasonably and is even not controled.

The cause of this problem is too large mtu value in "tc filter" command run by libvirt. The command uses burst value to set mtu and the burst is equal to average rate value if it's not set. This value is too large. For example if the average rate is set to 1024 kilobytes/s, the mtu value is set to 1024 kilobytes. That's too large compared to the size of network packets. Here libvirt applies tc ingress filter to Host's vnet(tun) device. Tc ingress filter is implemented with TBF(Token Buckets Filter) algorithm. TBF uses mtu value to calculate the amount of token consumed by each packet. With too large mtu value, the token consumption rate is set too large. This leads to token starvation and deterioration of TCP throughput.

Then, should we use the default mtu value 2 kilobytes? The anser is No, because Guest with virtio-net device uses 65536 bytes as mtu to transmit packets to Host, and the tc filter with the default mtu value 2k drops packets whose size is larger than 2k. So, the most packets is droped and again leads to deterioration of TCP throughput.

The appropriate mtu value is 65536 bytes which is equal to the maximum value of network interface device defined in <linux/netdevice.h>. The value is not so large that it causes token starvation and not so small that it drops most packets. Therefore this patch set the mtu value to 64kb(== 65535 bytes).

Again, here are the outbound QoS parameters and the TCP throughput with the libvirt patched.

outbound average rate[kilobytes/s] : Guest to Host throughput[Mbit/s] ====================================================================== 1024 (8Mbit/s) : 8.22 2048 (16Mbit/s) : 16.42 4096 (32Mbit/s) : 32.93 8192 (64Mbit/s) : 66.85 16384 (128Mbit/s) : 133.88 32768 (256Mbit/s) : 271.01 65536 (512Mbit/s) : 547.32

The outbound traffic conforms to the given limit.

Thank you,

Signed-off-by: Eiichi Tsukata <eiichi.tsukata.xh@hitachi.com> --- AUTHORS | 1 + src/util/virnetdevbandwidth.c | 2 +- 2 files changed, 2 insertions(+), 1 deletions(-)

diff --git a/AUTHORS b/AUTHORS index b876ee6..375db24 100644 --- a/AUTHORS +++ b/AUTHORS @@ -246,6 +246,7 @@ Patches have also been contributed by: Gerd Hoffmann <kraxel@redhat.com> Viktor Mihajlovski <mihajlov@linux.vnet.ibm.com> Thang Pham <thang.pham@us.ibm.com> + Eiichi Tsukata <eiichi.tsukata.xh@hitachi.com>

[....send patches to get your name here....]

diff --git a/src/util/virnetdevbandwidth.c b/src/util/virnetdevbandwidth.c index b9bd2e3..93df5c1 100644 --- a/src/util/virnetdevbandwidth.c +++ b/src/util/virnetdevbandwidth.c @@ -136,7 +136,7 @@ virNetDevBandwidthSet(const char *ifname, virCommandAddArgList(cmd, "filter", "add", "dev", ifname, "parent", "ffff:", "protocol", "ip", "u32", "match", "ip", "src", "0.0.0.0/0", "police", "rate", average, - "burst", burst, "mtu", burst, "drop", "flowid", + "burst", burst, "mtu", "64kb", "drop", "flowid", ":1", NULL);

if (virCommandRun(cmd, NULL) < 0)

Yeah, this tc magic was the hardest part when I've introduced QoS feature. I tested it on small values where this has not fully showed.

ACK and pushed. Thanks!

And special thanks for a very detailed explanation :-), excellent ! Daniel -- Daniel Veillard | libxml Gnome XML XSLT toolkit http://xmlsoft.org/ daniel@veillard.com | Rpmfind RPM search engine http://rpmfind.net/ http://veillard.com/ | virtualization library http://libvirt.org/