Re: CAP and performance problem

[George Dunlap <George.Dunlap@eu.citrix.com>]

On Tue, May 21, 2013 at 11:41 AM, George Dunlap
<George.Dunlap@eu.citrix.com> wrote:
> On Tue, May 21, 2013 at 10:02 AM, Massimo Canonico <mex@di.unipmn.it> wrote:
>> Hi,
>> I sent the following problem on xen-user ML without an answer. I hope I'll
>> get one in this ML.
>>
>> My application is written in std C++ and it makes a matrix multiplication:
>> so it uses only CPU and memory (no I/O, no network).
>>
>>
>> I'm quite surprise that with CAP = 100% I got my results in about 600
>> seconds and with CAP = 50% I got my results in about 1800 seconds (around 3
>> times longer).
>>
>> For this kind of application I was expecting to get results in about 1200
>> seconds (2 times longer) for the second scenario with respect to the first
>> one.
>>
>> Of course, the HW and SW are exactly the same for the 2 experiments.
>>
>> Am I wrong or the CAP mechanism is not working well?
>
> How many vcpus do you have?
>
> I'm not sure the CAP stuff actually does anything sensible with
> multiple cpus -- e.g., a VM with 2 vcpus maxed out will typically
> actually get 200%, so will a CAP force both of them to get only 25%
> each?
>
> The other possibility if you've got multiple vcpus is that you're
> running into something I call "concurrency hazard": if one vcpu grabs
> a spinlock and then is paused in order to enforce the cap, and another
> vcpu tries to grab that spinlock, it will just spin until the first
> vcpu manages to wake up and release it again.

Sorry, meant to add: if you have more than one vcpu and you want to
limit the cpu footprint, you're much better off halving the number of
vcpus rather than giving each vcpu only 50% of a cpu.

I gave a presentation on why here, if you're interested:

http://www.citrix.com/tv/#videos/2930

 -George