All of lore.kernel.org
 help / color / mirror / Atom feed
* LONG: An Asymmetric-Aware Scheduler for Hypervisor
       [not found] <CA+Y3Ckr-V2MiNCUFv4R4-uPzqFN+XS1d429kFpzJgGvbgJszZw@mail.gmail.com>
@ 2011-08-26 19:35 ` Nilesh Somani
  2011-08-26 19:44   ` Pasi Kärkkäinen
  2011-08-26 20:26   ` Kaushik Barde
  0 siblings, 2 replies; 3+ messages in thread
From: Nilesh Somani @ 2011-08-26 19:35 UTC (permalink / raw)
  To: xen-devel

*         Topic: *Description on How to make the scheduler of Hypervisor
asymmetric aware.

        The processor's have evolved from single-core to multi-core
processors. But these Processors
have been Symmetric ie identical.Every day the data on web in increasing and
the companies need
to set up large and powerful data center's. The power consumption is thus
increasing to a large extent.
Even as the requests on web increase the number of core's also need to
increased to speed up the
processing in large data center's. But with increasing number of cores the
power consumption is also
increasing.
        If the application has less Instruction level parallelism then its
latency on high core processor that
supports high level of parallelism increases with power consumption. So to
deal with these situations
came Asymmetric Multi core Processors.
       These are processor's with some fast cores and some slow cores. Now
the application requiring
low ILP can run on slow cores thus reducing the latency and the power
consumption. The area required
for 4 fast cores can have 2 fast cores and 6 small cores. The performance
increase in multithreading
is anytime more on 8 cores(2-fast and 6-slow) rather than 4 fast cores.
        What I came up in a paper is that the hypervisor's today are
asymmetric unaware. Thus for every
request it gets it treats all the processor's identical even in asymmetric
hardware. Consider IBM
Blade Server QS series server that has 2 cell processors of 3.2 GHz. This
means there are 2 fast cores
and 16 slow cores compared to the fast ones. It might happen that the
request to hypervisor is for fast
core but suppose both the fast cores are not idle and as the hypervisor
treats all the cores one and the same
it might transfer the request to slow core which is not desired.
       So to deal with such situations and to improve the efficiency in
power i propose to modify the scheduler
of the Xen hypervisor and make it asymmetric aware to schedule the jobs
properly ie in a manner to increase
the overall efficiency. Care also needs to be taken that the fast cores dont
go idle before the slow cores to
increase the efficiency and performance. Also there should be fairness among
the jobs requesting fast or
slow cores. We can also assign priorities to VM's such that the requests
from a particular VM goes to either
fast or slow cores.

I had some doubts regarding this.
I went through some IEEE and ASM papers and there has been proposals to make
schedulers aware of AMPs. So why hasn't it not yet being implemented in Xen?
Are there any issues?
Is there any other hardware other than Cell processor to go on with the
idea?

Nilesh Somani,
graduate student in Information technology
Contact:niel.doxford@gmail.com

^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: LONG: An Asymmetric-Aware Scheduler for Hypervisor
  2011-08-26 19:35 ` LONG: An Asymmetric-Aware Scheduler for Hypervisor Nilesh Somani
@ 2011-08-26 19:44   ` Pasi Kärkkäinen
  2011-08-26 20:26   ` Kaushik Barde
  1 sibling, 0 replies; 3+ messages in thread
From: Pasi Kärkkäinen @ 2011-08-26 19:44 UTC (permalink / raw)
  To: Nilesh Somani; +Cc: xen-devel

On Sat, Aug 27, 2011 at 01:05:55AM +0530, Nilesh Somani wrote:
> *         Topic: *Description on How to make the scheduler of Hypervisor
> asymmetric aware.
> 
>         The processor's have evolved from single-core to multi-core
> processors. But these Processors
> have been Symmetric ie identical.Every day the data on web in increasing and
> the companies need
> to set up large and powerful data center's. The power consumption is thus
> increasing to a large extent.
> Even as the requests on web increase the number of core's also need to
> increased to speed up the
> processing in large data center's. But with increasing number of cores the
> power consumption is also
> increasing.
>         If the application has less Instruction level parallelism then its
> latency on high core processor that
> supports high level of parallelism increases with power consumption. So to
> deal with these situations
> came Asymmetric Multi core Processors.
>        These are processor's with some fast cores and some slow cores. Now
> the application requiring
> low ILP can run on slow cores thus reducing the latency and the power
> consumption. The area required
> for 4 fast cores can have 2 fast cores and 6 small cores. The performance
> increase in multithreading
> is anytime more on 8 cores(2-fast and 6-slow) rather than 4 fast cores.
>         What I came up in a paper is that the hypervisor's today are
> asymmetric unaware. Thus for every
> request it gets it treats all the processor's identical even in asymmetric
> hardware. Consider IBM
> Blade Server QS series server that has 2 cell processors of 3.2 GHz. This
> means there are 2 fast cores
> and 16 slow cores compared to the fast ones. It might happen that the
> request to hypervisor is for fast
> core but suppose both the fast cores are not idle and as the hypervisor
> treats all the cores one and the same
> it might transfer the request to slow core which is not desired.
>        So to deal with such situations and to improve the efficiency in
> power i propose to modify the scheduler
> of the Xen hypervisor and make it asymmetric aware to schedule the jobs
> properly ie in a manner to increase
> the overall efficiency. Care also needs to be taken that the fast cores dont
> go idle before the slow cores to
> increase the efficiency and performance. Also there should be fairness among
> the jobs requesting fast or
> slow cores. We can also assign priorities to VM's such that the requests
> from a particular VM goes to either
> fast or slow cores.
> 
> I had some doubts regarding this.
> I went through some IEEE and ASM papers and there has been proposals to make
> schedulers aware of AMPs. So why hasn't it not yet being implemented in Xen?
> Are there any issues?
> Is there any other hardware other than Cell processor to go on with the
> idea?
> 

The first problem is that Xen hasn't been ported to PPC/Cell.
(http://wiki.xensource.com/xenwiki/XenPPC)

Are there Asymmetric x86/x64 systems out there?

-- Pasi

^ permalink raw reply	[flat|nested] 3+ messages in thread

* RE: LONG: An Asymmetric-Aware Scheduler for Hypervisor
  2011-08-26 19:35 ` LONG: An Asymmetric-Aware Scheduler for Hypervisor Nilesh Somani
  2011-08-26 19:44   ` Pasi Kärkkäinen
@ 2011-08-26 20:26   ` Kaushik Barde
  1 sibling, 0 replies; 3+ messages in thread
From: Kaushik Barde @ 2011-08-26 20:26 UTC (permalink / raw)
  To: 'Nilesh Somani', xen-devel

One simple reason is that Xen runs primarily a x86 systems, and x86's have
always been a symmetric core architecture (SMP). 
I suppose, since 2004 -- with Intel's La Grande/AMD's Pacifica - hardware
assists for Virtualization have been pretty much limited to x86, that is
changing with ARM A15.

However, my two cents-- I don't believe there should be any s/w limitation
in Xen or its system elements such as scheduler from running it on AMP
architectures. 

-Kaushik

-----Original Message-----
From: xen-devel-bounces@lists.xensource.com
[mailto:xen-devel-bounces@lists.xensource.com] On Behalf Of Nilesh Somani
Sent: Friday, August 26, 2011 12:36 PM
To: xen-devel@lists.xensource.com
Subject: [Xen-devel] LONG: An Asymmetric-Aware Scheduler for Hypervisor

*         Topic: *Description on How to make the scheduler of Hypervisor
asymmetric aware.

        The processor's have evolved from single-core to multi-core
processors. But these Processors
have been Symmetric ie identical.Every day the data on web in increasing and
the companies need
to set up large and powerful data center's. The power consumption is thus
increasing to a large extent.
Even as the requests on web increase the number of core's also need to
increased to speed up the
processing in large data center's. But with increasing number of cores the
power consumption is also
increasing.
        If the application has less Instruction level parallelism then its
latency on high core processor that
supports high level of parallelism increases with power consumption. So to
deal with these situations
came Asymmetric Multi core Processors.
       These are processor's with some fast cores and some slow cores. Now
the application requiring
low ILP can run on slow cores thus reducing the latency and the power
consumption. The area required
for 4 fast cores can have 2 fast cores and 6 small cores. The performance
increase in multithreading
is anytime more on 8 cores(2-fast and 6-slow) rather than 4 fast cores.
        What I came up in a paper is that the hypervisor's today are
asymmetric unaware. Thus for every
request it gets it treats all the processor's identical even in asymmetric
hardware. Consider IBM
Blade Server QS series server that has 2 cell processors of 3.2 GHz. This
means there are 2 fast cores
and 16 slow cores compared to the fast ones. It might happen that the
request to hypervisor is for fast
core but suppose both the fast cores are not idle and as the hypervisor
treats all the cores one and the same
it might transfer the request to slow core which is not desired.
       So to deal with such situations and to improve the efficiency in
power i propose to modify the scheduler
of the Xen hypervisor and make it asymmetric aware to schedule the jobs
properly ie in a manner to increase
the overall efficiency. Care also needs to be taken that the fast cores dont
go idle before the slow cores to
increase the efficiency and performance. Also there should be fairness among
the jobs requesting fast or
slow cores. We can also assign priorities to VM's such that the requests
from a particular VM goes to either
fast or slow cores.

I had some doubts regarding this.
I went through some IEEE and ASM papers and there has been proposals to make
schedulers aware of AMPs. So why hasn't it not yet being implemented in Xen?
Are there any issues?
Is there any other hardware other than Cell processor to go on with the
idea?

Nilesh Somani,
graduate student in Information technology
Contact:niel.doxford@gmail.com

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

^ permalink raw reply	[flat|nested] 3+ messages in thread

end of thread, other threads:[~2011-08-26 20:26 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
     [not found] <CA+Y3Ckr-V2MiNCUFv4R4-uPzqFN+XS1d429kFpzJgGvbgJszZw@mail.gmail.com>
2011-08-26 19:35 ` LONG: An Asymmetric-Aware Scheduler for Hypervisor Nilesh Somani
2011-08-26 19:44   ` Pasi Kärkkäinen
2011-08-26 20:26   ` Kaushik Barde

This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.