Xen users - Sep 2006 - xen scheduling

Hi all,

I have xen 3.0.2 installed on our server with 5+1 domains running. The 
CPU load is usually below 1 on all domains, but while 2 of them runs 
very well (including dom0), all others have extermely poor performance. 
After some research I found that these domains simply don''t receive 
enough CPU time - I wrote a small script which does some calculations 
and the results speak for themselves:

root@sql:~# time perl proba.pl
real    0m4.762s
user    0m4.724s
sys     0m0.000s

root@mail:~# time perl proba.pl
real    0m21.299s
user    0m0.168s
sys     0m0.004s
(I sent a Ctrl-C to 21 seconds)

I also found out that if I shutdown sql and then I start mail, mail will 
be the fast domain, so timing only depends on the sequence I start the 
domains (all domains have the same kernel).

I assumed that this should be a scheduling problem, so I tried to adjust 
the sched-sedf parameters. But no matter what I pass to xm sched-sedf, 
parameters remain the same:
root@evermind:/var/log# xm sched-sedf
Name                              ID Period(ms) Slice(ms) Lat(ms) Extra 
Weight
Domain-0                           0      20.0      15.0     0.0      
1      0
devel                             19     100.0       0.0     0.0      
1      0
mail                              16     100.0       0.0     0.0      
1      0
privateweb                        17     100.0       0.0     0.0      
1      0
publicweb                          5     100.0       0.0     0.0      
1      0
sql                               18     100.0       0.0     0.0      
1      0

I even tried to assign 2 vcpus to the slow host (I have a dual-core 
Opteron), but that didn''t help either.
Any ideas?

I''m using xen-hypervisor-3.0-i386-pae 3.0.2+hg9697-0bpo1 from the 
backports.org package with kernel 2.6.16.24, and of course, xen-utils 
with the same version.

Thanks,
gyim


_______________________________________________
Xen-users mailing list
Xen-users@lists.xensource.com
http://lists.xensource.com/xen-users

Hi,

Yesterday I rebooted the machine with option sched=bvt and now it seems 
to work perfectly. Did anybody else experience such issues with the sedf 
scheduler?

gyim

Gyimesi Ákos írta:
> Hi all,
>
> I have xen 3.0.2 installed on our server with 5+1 domains running. The 
> CPU load is usually below 1 on all domains, but while 2 of them runs 
> very well (including dom0), all others have extermely poor 
> performance. After some research I found that these domains simply 
> don''t receive enough CPU time - I wrote a small script which does
some
> calculations and the results speak for themselves:
>
> root@sql:~# time perl proba.pl
> real    0m4.762s
> user    0m4.724s
> sys     0m0.000s
>
> root@mail:~# time perl proba.pl
> real    0m21.299s
> user    0m0.168s
> sys     0m0.004s
> (I sent a Ctrl-C to 21 seconds)
>
> I also found out that if I shutdown sql and then I start mail, mail 
> will be the fast domain, so timing only depends on the sequence I 
> start the domains (all domains have the same kernel).
>
> I assumed that this should be a scheduling problem, so I tried to 
> adjust the sched-sedf parameters. But no matter what I pass to xm 
> sched-sedf, parameters remain the same:
> root@evermind:/var/log# xm sched-sedf
> Name                              ID Period(ms) Slice(ms) Lat(ms) 
> Extra Weight
> Domain-0                           0      20.0      15.0     0.0      
> 1      0
> devel                             19     100.0       0.0     0.0      
> 1      0
> mail                              16     100.0       0.0     0.0      
> 1      0
> privateweb                        17     100.0       0.0     0.0      
> 1      0
> publicweb                          5     100.0       0.0     0.0      
> 1      0
> sql                               18     100.0       0.0     0.0      
> 1      0
>
> I even tried to assign 2 vcpus to the slow host (I have a dual-core 
> Opteron), but that didn''t help either.
> Any ideas?
>
> I''m using xen-hypervisor-3.0-i386-pae 3.0.2+hg9697-0bpo1 from the 
> backports.org package with kernel 2.6.16.24, and of course, xen-utils 
> with the same version.
>
> Thanks,
> gyim
>
>
> _______________________________________________
> Xen-users mailing list
> Xen-users@lists.xensource.com
> http://lists.xensource.com/xen-users

_______________________________________________
Xen-users mailing list
Xen-users@lists.xensource.com
http://lists.xensource.com/xen-users