[Xen-users] sEDF Scheduling question

["Koen van Besien" <koen.vanbesien@xxxxxxxxx>]

Hi,

I was wondering if there is some more information about the sEDF
scheduler and its parameters.
Or people who have some experience with it.

The information I have found is from three papers concerning Xen.

* comparison of three cpu schedulers
www.hpl.hp.com/personal/Lucy_Cherkasova/papers/per-3sched-xen.pdf

* Xen and Co
csl.cse.psu.edu/publications/vee07.pdf

* Scheduling I/O in virtual machine monitors
www.cs.rice.edu/CS/Architecture/docs/ongaro-vee08.pdf


What I'm trying to do is to figure out what the best options are for
NFS performance and later on for webserver performance.
So I want to check what the influence is of the period and slice
allocation and also the impact of
being work conserving or not. ( the extra-parameter of sedf).
Later on I want to do this for the credit scheduler which should be
better on my dual core machine.

This is my machine setup:

Debian etch with 3 debian domUs all with LVM
Each has 256 MB RAM and dom0 has 1 GB of RAM.

o       AMD Athlon 64 X2 Dual Core 5000+
o       4 x 512 MB DDR ram: Total 2GB
o       Seagate 160GB S-ATA II 8MB
o       Debian Etch 64 bit / LVM / ext 3


I have 3 domUs running. One domU is running an nfs client , another
one the nfs server. Also dom0 is running an nfs server.
The third one is either idle or has a 100% cpu load via the stress
tool: http://weather.ou.edu/~apw/projects/stress/.

I write and read a file via the dd command on the NFS share which is
mounted via tcp
(mount -t nfs -o tcp 10.10.3.166:/home /mnt/helios166/)

write = dd if=/dev/zero of=./big.file bs=1M  count=500
read =  dd if=./big.file  of=/dev/null bs=1M


I shortly summarize my results:

First test ( period 10 ms work conserving)
========

* period of 10 ms work conserving
      xm sched-sedf 0 -p 10 -s 5 -l 0 -e 1 and this for all other
domains the same

* with the third vm running idle I get these results
     domU write = 35 sec
     domU read = 44 sec

     dom0 write = 16 sec
     dom0 read = 22 sec

* with the third vm running a 100% cpu load
     domU write = 31 sec
     domU read = 87 sec

     dom0 write = 19 sec
     dom0 read = 170 sec


second test  ( period 10 ms non work conserving)
========

* period of 10 ms NON work conserving
      xm sched-sedf 0 -p 10 -s 5 -l 0 -e 0 and this for all other
domains the same

* with the third vm running idle I get these results
     domU write = 104 sec
     domU read = 94 sec

     dom0 write = 59 sec
     dom0 read = 115 sec

* with the third vm running a 100% cpu load
     domU write = 110 sec
     domU read = 97 sec

     dom0 write = 95 sec
     dom0 read = 120 sec


third test ( period of 100 ms work conserving)
========

* period of 100 ms work conserving
      xm sched-sedf 0 -p 100 -s 50 -l 0 -e 1 and this for all other
domains the same

* with the third vm running idle I get these results
     domU write = 27 sec
     domU read = 86 sec

     dom0 write = 20 sec
     dom0 read = 14 sec

* with the third vm running a 100% cpu load
     domU write = 57 sec
     domU read = 74 sec

     dom0 write = 34 sec
     dom0 read = 30 sec

fourth test ( period of 100 ms NON work conserving)
========

* period of 100 ms NON work conserving
      xm sched-sedf 1 -p 100 -s 50 -l 0 -e 0 and this for all other
domains the same
      REMARK= setting dom0 to be NON work conserving gives weird
problems with it. Its constantly using 80% cpu

* with the third vm running idle I get these results
     domU write = 17 sec
     domU read = 12.8 sec

     dom0 write = 15.68 sec
     dom0 read = 17.5 sec

* with the third vm running a 100% cpu load
     domU write = 15.8 sec
     domU read = 16 sec

     dom0 write = 16.59 sec
     dom0 read = 13 sec


So concluding out of these results when you use NWC mode with a longer
period you get the
best results?

I it is quick testing and I just want to figure out what the exact
parameters are an the influence of them.
Are there some more benchmarks about this?

When I do a tcpdump on the nfs client I get a lot of duplicate acks
and TCP retransmission errors?
How is this possible as everything is virtual?

Are their some recommendation about which parameters to give to dom0 ?
I would guess that because
its not using a lot of cpu I should not given a lot of cpu share? Or
is it really important for I/O interrupt handling
that the dom0 is scheduled regularly?

I also don't know what the latency parameter is about?

So if people have some experiences about this or can give me some pointers


greetings

Koen

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