perfsonar development work

[Maciej Glowiak <>]

Hi Michael,

My previous e-mail should give you an idea what was the motivation for 
new psBase (except the performance) and how we would like to solve 
problems. Although we didn't analize the code so deeply. So, we're using 
"we know where the problem is" approach, however we made some 
performance testing (but not full profiling).

Now I am a bit busy with dLS implementation (using new base), but if you 
want to do some extra tests of old or new base, it would be great.

Maciej


Michael Bischoff wrote:
> Hello all,
>
> I've tried to harvest most of the idea's flying around and summarise them 
> here:
>
> Optimizing large data transfers (including Push Interface)
>
> * Leader: RL
> * Description: Return to project of push proposal. It should also cover the 
> possibility of
> sending measurement data using non-XML protocol
> * Expected goals:
>    o Identify benefits and problems
>    o Work on sequence diagram
>
> Inline encoded information in xml
> use two channels: control channel + data channel
>
> Changing xml handling:
> - Streaming DOM
> - SAX
> - STAX
>
> Perhaps use different strategies for xml-> Object model and Object -> xml for 
> example use dom
> for incoming and a streaming approach for outgoing.
>
> Optimise tomcat?
>
> I think it is important to quantify(instrumentation) and make a informed 
> decision based on
> that, we all have (converting) idea's about optimisation.
> (Not just the server-side but also the client.)
>
> To add my 2 cents: I think it's dangerous to presume things about what 
> optimisation would
> work; from practise we discovered that the worst hotspot is almost never 
> where you presume it
> is.
>
> To quote Jeff Kesselman(author of http://java.sun.com/docs/books/performance/)
> "You may think you know what your performance problem is, but you are 
> probably wrong. I've
> been tuning code the better part of 15 years and I don't think my initial 
> guess as to what my
> problem was has been right more then once or even close more then twice."
>
> On an other note implementing streaming would require quite a change to the 
> personar base
> code, this is just because how the data currently reaches the ServiceEngine. 
> Although there
> are elegant solutions around that too. esp with respect to 'incremental DOM'
>
> On jet an other note: I can't speak for the whole Java crowd but sending 
> serialised java
> around isn't something I would lean to Java RMI uses that and it gave 
> everyone performance
> headaches.
>
> About instrumenting/profiling
>
> For java:
> Netbeans has en excellent profiler
> Management beans (I think they where introduced in 1.5)
>
> In general, I think that ideally, we would want a continuous integration 
> server which also
> does profiling so we can monitor performance improvements along different 
> versions and also
> spot regressions right away.
>
> Best regards,
>
> Michael Bischoff


--

 --------------------------------------------------------------------
| Maciej Glowiak        Network Research and Development            ||
| 

     Poznan Supercomputing and Networking Center ||
| (+48 61) 858 2024          -- skype_id: maciej_psnc   GG: 4526858 ||
 ====================================================================