On 08/03/2014 08:16 PM, Nathan Scott wrote:
> Hey Will,
>
> ----- Original Message -----
>> Hi,
>>
>> Thursday evening I went to the a docker meetup session in Raleigh
>> (http://www.meetup.com/Docker-Raleigh/events/196434612/) where Jeremy Eder
>> talked about performance measurement and tuning for docker. He mentioned
>> using PCP in his slides
>> (http://slides.com/jeremyeder/docker-performance-analysis#/). He also
>
> I'm getting a 404 there, but found this via the meetup link:
> http://redhat.slides.com/jeremyeder/performance-analysis-of-docker
I might have munged the URL.
>
>> talked about using docker was being used to test scaling of Red Hat
>> Satellite/RHN. Docker was being used to provide a large number of clients
>> with lower overhead than using traditional KVM guests to stress test the the
>> server code. This approach could be useful for testing PCP to ensure that
>> PCP works well when monitoring large collections of machines.
>
> *nod* - great idea!
The ways of stressing pcp in the past have been to run a single pmcd daemon and
have multiple pmloggers trying to connect to it at the same time to get a feel
for what is going on. However, this model is a bit different to what would
likely be setup in clustered environments. In the clustered environment there
are going to be many machine each running a pmcd and a monitoring elsewhere
polling all the pmcd daemons for information. This migh provide a more
realistic stress on the monitor and avoid the single threaded bottleneck of
multiple monitors polling the same pmcd. There may be alternative way using
docker containers falls short, such as bottlenecks in the kernel such as procfs
showing up.
Alex Krzos at Red Hat did the work on Satellite/RHN testing. He has given me a
pointer to the code, but it is going to take me a little work to get up to
speed on how to use docker and how pcp test framework works.
-Will
|