[PVE-User] "fairsched" still in use with virtio guests in 1.4 (and slowness happen)

[Tomasz Chmielewski]

Tomasz Chmielewski wrote:
> Dietmar Maurer wrote:
>> There was small bug in the debian control file of qemu-server_1.1-7_amd64.deb (still contains version 1.1-6)
>>
>> Please download again and install with 'dpkg' if you see problems.
> 
> FYI - I still see a KVM guest which was makes lots of IO getting slow after a few days (best seen with long ping replies it sends):
> 
> 64 bytes from 192.168.111.173: icmp_seq=96 ttl=63 time=1.60 ms
> 64 bytes from 192.168.111.173: icmp_seq=97 ttl=63 time=1.90 ms
> 64 bytes from 192.168.111.173: icmp_seq=98 ttl=63 time=563 ms
> 64 bytes from 192.168.111.173: icmp_seq=99 ttl=63 time=1.64 ms
> 64 bytes from 192.168.111.173: icmp_seq=100 ttl=63 time=998 ms
> 64 bytes from 192.168.111.173: icmp_seq=101 ttl=63 time=1.56 ms
> 
> 
> Although it is an order of magnitude longer until it gets into that state now.
> 
> No idea how to debug it farther, but for now, I'd recommend using e1000 for guests making extensive CPU/IO usage.

There must be something fishy in the 2.6.24/OpenVZ kernel.

~3 weeks ago, I've compiled a vanilla 2.6.31.6 kernel.

KVM guests doing intensive IO (disk + net) work fine, no "slowness" 
after some time.


I can only conclude that:
- either 2.6.24 kernel, or OpenVZ adds some incompatibility here,
- one should not use virtio_net, but e1000 instead


-- 
Tomasz Chmielewski
http://wpkg.org