RE: Performance with multiple users during transformat

Hi,

> Thanks for the information. So are you firing off
> the requests manually,
> simply by hitting the send key simultaneously on
> three browsers?
Yes, you are right :-)

> If four transformations run concurrently, with no
> contention between them,
> then you would expect each one to take four times as
> long as when the
> processor is dedicated to a single user. This seems
> to be consistent with
> the results you observe.

Does this mean that all the four requests are being
processed by a single processor. Are the XSLT
processors internally single threaded ?
If not then why should T*N come into picture when we
have a huge UNIX box and only 4 testers are using the
application now.


>Achieving  optimal response time then
> depends on scheduling the work intelligently to
> avoid overloading the
> available memory, which depends on how you configure
> your application
> server. The ideal concurrency in this environment is
> probably obtained by
> taking the available memory, subtracting the shared
> memory requirement, and
> dividing what's left by the memory requirement of
> each transformation.

Let me play around with the JVM / Heap size /
Websphere setting a bit and I will publish the results
to the group again.

Thanks,
Bhupendra.




 --- Michael Kay <mike@xxxxxxxxxxxx> wrote: 
> Thanks for the information. So are you firing off
> the requests manually,
> simply by hitting the send key simultaneously on
> three browsers?
> 
> If four transformations run concurrently, with no
> contention between them,
> then you would expect each one to take four times as
> long as when the
> processor is dedicated to a single user. This seems
> to be consistent with
> the results you observe.
> 
> In a real environment the arrival time of requests
> will be randomly
> distributed, and there will be significant think
> time, so the response time
> will be much better than T*N where T is the
> transformation time and N is the
> number of users. 
> 
> As the number of users increases you have a classic
> transaction processing
> scenario. The theoretical maximum throughput you can
> achieve
> (transformations per second) is determined by
> dividing the machine capacity
> by the cost of each transformation. Achieving
> optimal response time then
> depends on scheduling the work intelligently to
> avoid overloading the
> available memory, which depends on how you configure
> your application
> server. The ideal concurrency in this environment is
> probably obtained by
> taking the available memory, subtracting the shared
> memory requirement, and
> dividing what's left by the memory requirement of
> each transformation. Don't
> ask me to translate this into concrete Websphere
> terms!
> 
> Michael Kay
> 
> > -----Original Message-----
> > From: Bhupendra Singh
> [mailto:skbhupendra@xxxxxxxxxxx] 
> > Sent: 01 December 2004 18:18
> > To: xsl-list@xxxxxxxxxxxxxxxxxxxxxx
> > Subject: RE:  Performance with multiple users
> during 
> > transformation
> > 
> > Michael ,
> > Thanks for the input, Heres what you requested
> for.
> > 
> > > You need to explain your measurement scenario
> more
> > > carefully. How is the
> > > XSLT processor running - as a servlet? 
> > 
> > The XSLT processor is running as a servlet. Its a
> Web
> > based application, which takes client request with
> > Servlets. The serlvet brings the data in XML
> format
> > and then uses the transformer to make another XML,
> > which is send back to the user.
> > 
> > >How are you
> > > measuring response time?
> > 
> > I am taking the diference of the time before and
> after
> > the Transformer.transform() method and printing it
> in
> > a log file.
> > 
> > > Are
> > > these transformations all
> > > using the same stylesheet? 
> > 
> > All the users are requesting the same page with
> the
> > same parameter, so the same input XML and
> stylesheet
> > is used.
> > 
> > >Are you saving the
> > > compiled stylesheet in memory?
> > 
> > I am storing the compiled XSLT as templates in a
> > Hashmap. I use this Template to create a new
> > Transformer object using the
> > templates.newTransformer().
> > Before taking the test I am running a dummy user
> > request to build this template in the Hashmap.
> > 
> > > Are they all running in the same thread, or do
> you
> > > have a pool of
> > > threads/processes?
> > Each user fires the request from different IE
> > sessions. Thus the servlet threads will be
> different
> > for each.
> > 
> > 
> > >how much memory you
> > > are allocating to the Java VM, 
> > I have kept the minimum JVM heap size to 1 GB for
> > Websphere App server.
> > 
> > 
> >  --- Michael Kay <mike@xxxxxxxxxxxx> wrote: 
> > > > I have a 250 KB XML to be transformed into
> another
> > > XMl
> > > > document. I am using a high end UNIX box with
> 8
> > > CPUs
> > > > and 32 Gig RAM. 
> > > > Intially I was using Xalan and the timings for
> > > > transformation are as follows:
> > > > 1 user : 3 secs
> > > > 2 users: 6 secs
> > > > 4 users: 8 secs
> > > > 
> > > > Then I changed to Saxon8 and the timings are
> > > > 1 user : 1.5 secs
> > > > 2 users: 3 secs
> > > > 4 users: 4 secs
> > > > (I fear whats going to happen when 1000 users
> will
> > > use
> > > > the applicaiton)
> > > 
> > > You need to explain your measurement scenario
> more
> > > carefully. How is the
> > > XSLT processor running - as a servlet? How are
> you
> > > measuring response time?
> > > What is the think time between transactions? Are
> > > these transformations all
> > > using the same stylesheet? Are you saving the
> > > compiled stylesheet in memory?
> > > Are they all running in the same thread, or do
> you
> > > have a pool of
> > > threads/processes?
> > > > 
> > > > Although I got a 50% performance enhancement
> with
> > > > Saxon for each scenario, but I still dont
> > > understand
> > > > one thing why do we have the increase in
> > > > transformation timing as the users increase.
> > > 
> > > That depends very much on how the work is being
> > > scheduled, which is why I
> > > asked about your execution environment. 
> > > 
> > > > Since Saxon is based on SAX parsing there
> should
> > > be no
> > > > memory constrained (Also my 32 Gig RAM is
> never
> > > > utilised more than 5GB at any time.)
> > > 
> > > No, that's a bad misunderstanding. Every XSLT
> > > processor today builds the
> > > source document in memory. However, 250Kb is not
> > > particularly big. Again,
> > > however, it depends on how you are scheduling
> the
> > > work, how much memory you
> > > are allocating to the Java VM, and so on. 
> > > > 
> > > > Can somebody briefly tell me how the
> > > transformation
> > > > works, starting from how the XSL is parserd,
> how
> > > the
> > > > XML is parsed and then the transformation.
> > > > Can the parser and the XSLT processors be
> > > different.
> > > 
> > > If you're using Saxon, then the article at
> > > http://www.xml.com/pub/r/1065 may
> > > be helpful.
> > > 
> > > Michael Kay
> > > http://www.saxonica.com/