On Thu, 2011-06-16 at 19:49 +0300, Apollon Oikonomopoulos wrote: > Package: linux-2.6 > Version: 2.6.32-34squeeze1 > Severity: normal > Tags: patch > > Hi, > > While performing stress testing on a KVM instance with IPv6, we noticed very > poor performance on IPv6 TCP connections, which basically came down to the bug > fixed by upstream commit 8e1e8a4779cb23c1d9f51e9223795e07ec54d77a[1]. Both, > the > host and the guest were running 2.6.32-5-amd64 2.6.32-31 and the host featured > Intel NICs, using e1000e and igb. SSH sessions to the guest from outside would > effectively stall when large output was generated (e.g. ls -lR /) and > the problem would disappear when disabling virtio_net's gso (modprobe > virtio_net gso=0) or disabling the host's TSO using ethtool. > > Furthermore, while trying to reproduce the behaviour on another system, we > triggered the bug fixed by commit 0aa68271510ae2b221d4b60892103837be63afe4[2], > i.e. a routed (not bridged) KVM instance using virtio-net would not get its > GSO > IPv6 packets forwarded to the network, with the same poor performance as > above.
David, these look like good candidates for longterm updates. What do you think? Ben. > After backporting and applying both patches to squeeze's version, IPv6 > performance was restored to the same levels as IPv4 in both cases. You > can find the backported versions of the patches attached. > > Thanks, > Apollon > > [1] > http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=8e1e8a4779cb23c1d9f51e9223795e07ec54d77a > [2] > http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=0aa68271510ae2b221d4b60892103837be63afe4 > > > -- Package-specific info: > ** Version: > Linux version 2.6.32-5-amd64 (Debian 2.6.32-31) (b...@decadent.org.uk) (gcc > version 4.3.5 (Debian 4.3.5-4) ) #1 SMP Mon Mar 7 21:35:22 UTC 2011 > > ** Command line: > BOOT_IMAGE=/vmlinuz-2.6.32-5-amd64 root=/dev/mapper/linux-root ro > > ** Not tainted -- Ben Hutchings I'm always amazed by the number of people who take up solipsism because they heard someone else explain it. - E*Borg on alt.fan.pratchett
signature.asc
Description: This is a digitally signed message part
