Open Stack

On 12/16/2013 02:25 AM, Neependra Khare wrote:
> Hi Jay,
>
> Thanks for your comments.  Please find my reply in-line.
>
> On 12/14/2013 12:58 AM, Jay Pipes wrote:
>> I have listed down the methodology I'll be following for this test:-
>>> https://wiki.openstack.org/wiki/KeystonePerformance#Identify_CPU.2C_Disk.2C_Memory.2C_Database_bottlenecks
>>>
>>
>> My first suggestion would be to rework the performance benchmarking
>> work items to have clearer indications regarding *what are the metrics
>> being tested* in each work item.
> Performance characterization is an iterative process. I am open to
> rework on the work-items as we
> go along.

Right, but the smaller the work item, the easier the iterations are :)

>> For example, the first work item is "Identify CPU, Disk, Memory, and
>> Database Bottlenecks".
>>
>> The first test case listed is:
>>
>> "Test #1, Create users in parallel and look for CPU, disk or memory
>> bottleneck."
>>
>> I think that is a bit too big of an initial bite ;)
>>
>> Instead, it may be more effective to instead break down the
>> performance analysis based on the metrics you wish to test and the
>> relative conclusions you wish your work to generate.
>>
>> For example, consider this possible work item:
>>
>> "Determine the maximum number of token authentication calls that can
>> be performed"
> Tests like these would be very subjective to the hardware and software
> resources we have like
> no. of CPUs, Memcahced etc.  Its is very important to see if we can find
> any obvious bottlenecks.

No, that's not my point. When you have a specific metric like "number of 
token authentication calls that can be performed in X minutes", you can 
iterate based on singular changes -- not to the hardware, but to the 
configuration of the software. If you are trying to solve the problem of 
"where are my bottlenecks", without first identifying what metrics will 
describe how a piece of software scales, then you are putting the cart 
before the horse.

>> Within that work item, you can then further expand a testing matrix,
>> like so:
>>
>> * Measure the total number of token authentication calls performed by
>> a single client against a single-process, Python-only Keystone server
>> * Measure the total number of token authentication calls performed by
>> a single client against a multi-process Keystone server running inside
>> an nginx or Apache container server -- with 2, 4, 8, 16, and 32
>> pre-forked processes
> Any pointers on configuring multi-process Keystone would be helpful. I
> see a method
> mentioned in "Run N keystone Processes" section of following:-
> http://blog.gridcentric.com/bid/318277/Boosting-OpenStack-s-Parallel-Performance"

Absolutely. You can spawn Keystone server in multiple pre-forked Apache 
processes by configuring Keystone in an Apache vhost. Some general docs:

http://docs.openstack.org/developer/keystone/apache-httpd.html

Take a look at provision.sh script in eNovance's keystone-wsgi-apache repo:

https://github.com/enovance/keystone-wsgi-apache/blob/master/provision.sh#L152

>> * Measure the above using increasing numbers of concurrent clients --
>> 10, 50, 100, 500, 1000.
>>
>> There's, of course, nothing wrong with measuring things like CPU, disk
>> and I/O performance during tests, however there should be a clear
>> metric that is being measured for each test.
> Agreed. Let me start collecting results from the tests you suggested
> above and I mentioned
> on the wiki. Once we have those, we can rework on the work-items. Does
> that sound OK ?

Sure, absolutely. I'm just big on first defining the metrics by which 
scale can be described, and THEN describing the test variations and 
iterations...

>> My second suggestion would be to drop the requirement of using RDO --
>> or any version of OpenStack for that matter.
> My end goal would be to have scripts that one can run on any of the
> OpenStack distribution.
> RDO is mentioned here an example here.

Probably worth just removing the RDO reference entirely from the wiki, 
since, as you agree below, benchmarking Keystone actually does not 
require installing OpenStack as a whole at all...

Best,
-jay

>> In these kinds of tests, where you are not measuring the integrated
>> performance of multiple endpoints, but are instead measuring the
>> performance of a single endpoint (Keystone), there's no reason, IMHO,
>> to install all of OpenStack. Installing and serving the Keystone
>> server (and it's various drivers) is all that is needed. The fewer
>> "balls up in the air" during a benchmarking session, the fewer
>> side-effects are around to effect the outcome of the benchmark...
> Agreed. As mentioned in following I suggested to install just Keystone
> on the instances, where the tests would be performed :-
> https://wiki.openstack.org/wiki/KeystonePerformance#Test_.231.2C_Create_users_in_parallel_and_look_for_CPU.2C_disk_or_memory_bottleneck.
>
>
> Thanks,
> Neependra
>
>
>>
>> Best,
>> -jay
>>
>>
>>
>> _______________________________________________
>> Mailing list:
>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack
>> Post to     : openstack at lists.openstack.org
>> Unsubscribe :
>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack
>