Re: [Xen-devel] [PATCH 09 of 10 [RFC]] xl: Introduce Best and Worst Fit guest placement algorithms

[Dario Faggioli <raistlin@xxxxxxxx>]

On Wed, 2012-04-11 at 15:17 +0200, Dario Faggioli wrote: 
> Besides than "auto" and "ffit", as per the previous patch, enable
> Best and Worst Fit placement heuristics to `xl`, for the user to
> chose them in the config file. Implementation just sits on top
> of the First Fit algorithm code, with just the proper reordering
> of the nodes and their free memory before invoking it (and after
> it finishes).
> 
> TODO: * As usual for this series, benchmarking and testing,
>         as much thorough and comprehensive as possible!
> 
I manage in getting some numbers from these placement heuristics too.

http://xenbits.xen.org/people/dariof/benchmarks/specjbb2005-numa_aggr/index.html

As a foreword, they come from some of the benchmarks I already run, and
from running some new ones, but with the same logic and infrastructure,
meaning:
- all the VMs were equal in (memory) size
- all the VMs had the same # of VCPUs
- all the VMs are always created in the very same order
These are definitely not the best condition for stressing the
algorithms, neither they are for trying to understand the algorithms'
differences and capabilities of bringing performance benefits... Anyway,
they're _something_, and a slight improvement wrt default placement is
already noticeable in the graphs. :-)

Look, for example, at one here this:

http://xenbits.xen.org/people/dariof/benchmarks/specjbb2005-numa_aggr/2.html

And compare the red curve (default placement/schedulig) with all the
other ones. They all represents the aggregate throughput achieved by the
SpecJBB 2005 benchmark --counting (and averaging it on) all the VMs--
when varying the number of VMs, so as usual for SpecJBB, higher is
better.

"wfit" and "bfit" clearly works better than "default", which would have
been desirable, and actually expected, especially for Worst Fit, in this
particular scenario. It is also normal that "ffit" does a worse job than
"default" until the number of VMs excedes 4. In fact, considering the
amount of RAM each VM has (without forgetting some RAM is assigned to
Dom0 as well), with less than 5 VMs all of them are fitted on the first
NUMA node, thus the memory accesses are local, but the load distribution
is clearly sub-optimal!

As for the scheduling related results, more serious testing and tracing
is needed in order to verify things are happening as expected, but the
graphs seemed nice to me, and I thus thought it could be worth sending
them, while working on having more interesting ones. :-)

Thanks and Regards,
Dario

-- 
<<This happens because I choose it to happen!>> (Raistlin Majere)
-----------------------------------------------------------------
Dario Faggioli, Ph.D, http://retis.sssup.it/people/faggioli
Senior Software Engineer, Citrix Systems R&D Ltd., Cambridge (UK)

Attachment: signature.asc
Description: This is a digitally signed message part

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel