sedf testing: volunteers please

sedf testing: volunteers please

[Stephan Diestelhorst]

The new sedf scheduler has been in the xen-unstable reopository for a
couple of days now. As it may become the default scheduler soon, any
testing now is much appreciated!

Quick summary can be found in docs/misc/sedf_scheduler_mini-HOWTO.txt

Future directions:
-effective scheduling of SMP-guests
  -clever SMP locking in domains (on the way)
  -timeslice donating (under construction)
  -identifying gangs and schedule them together
  -balancing of domains/ VCPUs

Any comments/wishes/ideas/... on that are welcome!

Best,
  Stephan Diestelhorst

Re: sedf testing: volunteers please

[Ryan Harper]

* Stephan Diestelhorst <sd386@cl.cam.ac.uk> [2005-05-18 09:04]:
> The new sedf scheduler has been in the xen-unstable reopository for a
> couple of days now. As it may become the default scheduler soon, any
> testing now is much appreciated!

I'll start testing edf out as my default scheduler.

> Quick summary can be found in docs/misc/sedf_scheduler_mini-HOWTO.txt
> 
> Future directions:
> -effective scheduling of SMP-guests
>   -clever SMP locking in domains (on the way)
>   -timeslice donating (under construction)

As you've seen in my post about forced vcpu selection, I'm very
interested in this.  Do you have any code that works with edf that I can
test out?

-- 
Ryan Harper
Software Engineer; Linux Technology Center
IBM Corp., Austin, Tx
(512) 838-9253   T/L: 678-9253
ryanh@us.ibm.com

Re: sedf testing: volunteers please

[Xuehai Zhang]

Stephan,

I enabled the sedf scheduler by applying the patch to Xen testing tree not
the unstable tree.

Then I did the following test. I started two user domains (named "vm1" and
"vm2" respectively). I made the following sedf configurations:

xm sedf vm1 0 0 0 0 2
xm sedf vm2 0 0 0 0 8

My intention is to have vm1 reserve 20% of the available cpu and vm2
reserve the rest of 80% (please correct me if my understanding about
sedf here is wrong).

Then I start "slurp" job in both domains and it will print out the cpu
share continuously. To my surprise, vm1 takes around 4% of cpu and vm2
occpuies around 17% cpu. I was expecting they share the cpu something like
20% and 80% though the ratio of 4% and 17% is similar as that of 20% and
80%. BTW, dom0 didn't run any extra job when I ran the test.

Could you please let me know why only 21% (4%+17%) cpu is given to both
vm1 and vm2 not 100%-% taken by dom0?

Thanks.

Xuehai

On Wed, 18 May 2005, Stephan Diestelhorst wrote:

> The new sedf scheduler has been in the xen-unstable reopository for a
> couple of days now. As it may become the default scheduler soon, any
> testing now is much appreciated!
>
> Quick summary can be found in docs/misc/sedf_scheduler_mini-HOWTO.txt
>
> Future directions:
> -effective scheduling of SMP-guests
>   -clever SMP locking in domains (on the way)
>   -timeslice donating (under construction)
>   -identifying gangs and schedule them together
>   -balancing of domains/ VCPUs
>
> Any comments/wishes/ideas/... on that are welcome!
>
> Best,
>   Stephan Diestelhorst
>
>
>
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xensource.com
> http://lists.xensource.com/xen-devel
>
>

Re: sedf testing: volunteers please

[Diwaker Gupta]

On 5/18/05, Stephan Diestelhorst <sd386@cl.cam.ac.uk> wrote:
> The new sedf scheduler has been in the xen-unstable reopository for a
> couple of days now. As it may become the default scheduler soon, any
> testing now is much appreciated!

I'm curious (sorry if this has already been discussed on the list, I
couldn't find it in a google search). Note that I'm not defending
either BVT or SEDF, I just wanted to know the reasons for the switch
from BVT as default :)

-- 
Diwaker Gupta
http://resolute.ucsd.edu/diwaker

Re: sedf testing: volunteers please

[xuehai zhang]

AFAIK, most features of BVT (proportinal CPU share based on weights especially) are included in SEDF 
except the time warping feature. But SEDF can also enforce abosolute CPU share for each domains via 
period/slice mechnism, which is similar as that in the broken atropos scheduler. This is a kind of 
hard guarantee which BVT can't provide.

Xuehai

Diwaker Gupta wrote:
> On 5/18/05, Stephan Diestelhorst <sd386@cl.cam.ac.uk> wrote:
> 
>>The new sedf scheduler has been in the xen-unstable reopository for a
>>couple of days now. As it may become the default scheduler soon, any
>>testing now is much appreciated!
> 
> 
> I'm curious (sorry if this has already been discussed on the list, I
> couldn't find it in a google search). Note that I'm not defending
> either BVT or SEDF, I just wanted to know the reasons for the switch
> from BVT as default :)
> 

Re: sedf testing: volunteers please

[Stephan Diestelhorst]

Xuehai,

  your idea is about right. Things to notice are:
How much slice did you give to dom0?
If dom0 gets 75% of the cpu, then the other two domains will share the
remaining 20%
100%-75% - 5% (a small reserve)) in the requsted ratio.
That means they will get 2/10 * 20% = 4% and 8/10 * 20% = 16% of the CPU.
This is guranteed and they won't exceed their reservation because they
don't have the extra-flag specified!
You could either:
  -reduce the reservation fordom0
or
  -use the extraflag for your domains vm1 and vm2 to givethem any
remaining time (which drives them in weighted extratime mode)
  xm sedf vm1 0 0 0 1 2

Hope that helps,
  Stephan

>Stephan,
>
>I enabled the sedf scheduler by applying the patch to Xen testing tree not
>the unstable tree.
>
>Then I did the following test. I started two user domains (named "vm1" and
>"vm2" respectively). I made the following sedf configurations:
>
>xm sedf vm1 0 0 0 0 2
>xm sedf vm2 0 0 0 0 8
>
>My intention is to have vm1 reserve 20% of the available cpu and vm2
>reserve the rest of 80% (please correct me if my understanding about
>sedf here is wrong).
>
>Then I start "slurp" job in both domains and it will print out the cpu
>share continuously. To my surprise, vm1 takes around 4% of cpu and vm2
>occpuies around 17% cpu. I was expecting they share the cpu something like
>20% and 80% though the ratio of 4% and 17% is similar as that of 20% and
>80%. BTW, dom0 didn't run any extra job when I ran the test.
>
>Could you please let me know why only 21% (4%+17%) cpu is given to both
>vm1 and vm2 not 100%-% taken by dom0?
>
>Thanks.
>
>Xuehai
>
>On Wed, 18 May 2005, Stephan Diestelhorst wrote:
>
>  
>
>>The new sedf scheduler has been in the xen-unstable reopository for a
>>couple of days now. As it may become the default scheduler soon, any
>>testing now is much appreciated!
>>
>>Quick summary can be found in docs/misc/sedf_scheduler_mini-HOWTO.txt
>>
>>Future directions:
>>-effective scheduling of SMP-guests
>>  -clever SMP locking in domains (on the way)
>>  -timeslice donating (under construction)
>>  -identifying gangs and schedule them together
>>  -balancing of domains/ VCPUs
>>
>>Any comments/wishes/ideas/... on that are welcome!
>>
>>Best,
>>  Stephan Diestelhorst
>>
>>
>>
>>_______________________________________________
>>Xen-devel mailing list
>>Xen-devel@lists.xensource.com
>>http://lists.xensource.com/xen-devel
>>
>>
>>    
>>