perfsonar development work

[Jason Zurawski <>]

Slawomir Trzaszczka wrote:
> On Mon, 2009-11-30 at 09:52 -0500, Jason Zurawski wrote:
> > Slawomir Trzaszczka wrote:
> > > Hi all,
> > >
> > > I've created performance tests for LS1 and LS2. In these tests, I used
> > > larger "size" of services then in  previous version of tests - services
> > > contain more interfaces(100,500,1000).
> > Couple of comments.  All in all good job.
> >
> > 1) 1000 isn't really the largest amount of interfaces we have seen in 
> > the wild.  Perhaps someone from ESnet can provide their store file for 
> > reference.
> >
> > 2) Also on the subject of interfaces, have you tried the case of 
> > 'updating' a registration?  This would be similar to a back to back 
> > de-registration and registration to simulate the change of a couple of 
> > interfaces.
>
> No. I will perform this kind of tests
>
> > 3) 'De-registration' by itself may not be a useful thing in the pS 
> > world.  I think it is a rarity that services will send a de-registration 
> > message before they go offline cleanly, and they will not send if they 
> > are go offline un-cleanly.  I think its good to test, but testing the 
> > timeout/cleaning of the database during the regular processing of 
> > registrations/queries may be more valuable.
>
> Ok, but what about cleanup action ? Cleanup is very similar to
> de-registration but it's internal action of LS.  


This depends on how the de-registration is implemented.  The most 
expensive part of the pS-PS LS de-registration is the query to find what 
must be removed.  The cleaning operation in our case is really just a 
linear search of the control metadata + the removal of each item that 
qualifies.

I think that coupling the cleaning (or summarization) with other 
operations such as a complex query or registration would be a good to 
think about as you modify the procedure.

-jason


> > 4) It was mentioned in previous mails, but serial testing is not as 
> > valuable as concurrent.  This will show what sort of bottleneck can 
> > occur if you are running two complex things at the same time.  This our 
> > biggest problem with DBXML in the pS-PS hLS/gLS.
>
> Ok, I'll prepare some tests. 
>
> > -jason
>
> Regards,
>
> Slawek