perfsonar development work

["Vedrin Jeliazkov" <>]

Hi,

Roman Lapacz 
<>
 wrote:

> Jeff W. Boote wrote:
> > Roman Lapacz wrote:
> >> Unfortunately the release 1.0 does not contain recent changes, which 
> >> I hope, improve the performance. It would be interesting to see the 
> >> tests of RRD MA from svn trunk or the next release (question to the 
> >> release team: is there a timetable for this?).
> >
> > Roman,
> >
> > What changes have you made to the RRD MA with respect to performance 
> > since the 1.0 release?
> 
> 
> - use of http type of access to xml db (Maciej's recent implementation; 
> I remember he sent to the mailing list the results of performance tests 
> of it)
> - when fetching from the metadata configuration file (getting the key) 
> only one xquery is used (previously 2 xpath statements were sent; so now 
> we've got only one connection to xml db instead of two)
> - when xml db is not configured as a storage the metadata configuration 
> file is parsed only once and kept in memory as an object (this object 
> can be used for fetching data using xquery/xpath)

I've run my usual tests against the latest RRD MA (SVN version from 2006-08-25
19:29 EEST) and have compared the results with those for perfSONAR-1.0. Here
are the timings I've got:

1. One MetadataKeyRequest (summary) [sec]       
2. Number of Interfaces (NoI)   
3. (NoI x 2) SetupDataRequests (detailed summary) [sec]
4. Average SetupDataRequest [msec]
5. Average Link latency (RTT) [msec]
6. Implementation
        
                (1)     (2)     (3)     (4)     (5)     (6)     
        
ISTF-J           14      78      48      308    0.204   perfSONAR-1.0
ISTF-J            4      78     537     3442    0.204   svn (HTTP)
ISTF-J            9      78     596     3821    0.204   svn (XMLRPC)

In all 3 cases I've run the tests twice and recorded the timings during the
second run in an attempt to avoid capturing various initialisations, which
occur only once.

Conclusions:

(1) MetadataKeyRequests are processed up to 3.5 times faster by the SVN
version, when HTTP access to eXist is configured;

(2) HTTP access to eXist provides better performance indeed (both for
MetadataKeyRequests and SetupDataRequests), when compared to XMLRPC;

(3) SetupDataRequest processing in the SVN version is more than 10 times
slower, compared to perfSONAR-1.0;

I'm not sure whether the SetupDataRequest slow down is due to local
configuration issues, forgotten debugging options or something else. It would
be nice to have the same (or even better) SetupDataRequest performance in the
SVN version.

Kind regards,
Vedrin