On Tuesday 21 November 2006 20:36, Michael Edwards wrote:
> Comments inline:
>
> On 11/21/06, Saravana Kumar <[EMAIL PROTECTED]> wrote:
> > On Monday 20 November 2006 20:01, Michael Edwards wrote:
> > Job ID               Username Queue    Jobname    SessID NDS   TSK Memory
> > Time S Time
> > -------------------- -------- -------- ---------- ------ ----- --- ------
> > ----- - -----
> > 66.hulk.gai.net      user workq    TestJob      6571     3   1    -- 
> > 10000 R 00:02
> >    node2+node1+node3
>
> That output only says that you have successfully reserved 3 nodes.
> That is, that job has the resources of node 1 2 and 3 available.
> Torque does not know if they are being used at all, that is something
> you would have to check with ganglia.  Open a web browser and go to
> http://localhost/ganglia and you will get a lot of performance
> information about the cluster.  Run the job and watch the cpu graphs
> of nodes 1,2 and 3.  I strongly suspect you will only see one of them
> do anything significant.
Yes. You are right. The code is using only one node's cpu. While other nodes 
are reserved but idle(not used).
>
> > I don't know why the queue(hp) that i created doesn't execute the job in
> > all hosts.
> >
> > > Do  you want one identical copy of the sas script to run on each node?
> >
> > I don't understand your question. I want the sas job to run in one
> > node(to keep file io in one node) but use all the free nodes' cpu cycles
> > which i think will improve the performance and decreases the code running
> > time.
>
> Unless SAS is much smarter than I think it is, this is unlikely to
> happen with out a lot of work on your part.  Clusters (of any sort,
> this is not an OSCAR problem) do not divide up work automatically or
> smartly as you (and lots of other people, don't feel bad) want them
> to.  The way programs work is too different for this to be done in any
> sort of intelegent way I have heard of.  There are some cluster
> programs (openMosix, kerrighed) which will try to do this to a very
> limited degree, but they still require very specially designed
> programs to gain any significant advantage.  Kerrighed has an Oscar
> project called SIS Oscar which is working on a 5.0 based version, I
> believe, if you want to try it.
Will try this when i have spare time.
>
> Anyway, the take home point is that to spread across nodes in the way
> you are describing requires specially writen software, which is very
> rare.  There are basically two ways clusters are used.  Custom writen
> (usually) code which uses MPI or some other communication protocol to
> split up the work among the nodes in a smart (custom) way.  This is
> called parrallel programing.  The other way is to have many
> non-parrallel jobs running on the cluster at the same time. You could
> run your SAS script on all 3 nodes for example, walking a parameter
> space.  This is called serial programing, and still usually requires
> some clever scripting to deal with the peculiarities of cluster
> computing.
>
> I am neglecting the type of automatic parallelizaton which you are
> describing because I see it as a special case of parallel programing
> which, as I mentioned, basically doesn't exist.
>
> Others may disagree with me on this point, so we'll see if they pipe in.
>
> > My sas jobs have lot of io. Is it better to run the sas job in one node
> > and use other node's cpu cycles or can the sasjob itself be split to
> > multiple disks and cpus with the script(not clear on this).
> >
> > The whole point of installing this cluster suite was to minimize the time
> > taken by sas jobs.
>
> I'll poke around and see if SAS has a parallel programing module, I
> know matlab does, but they still would require extra coding to use.
>
> The cluster can still help improve your overall processing time, but
> it will probably be in a serial mode rather than a parallel mode.
I wanted to implement the cluster to overcome the number of jobs run in 
parallel and cause iowait issues in my servers.

May be i must look out for solutions to increase the io speed(clustered file 
systems like coda/ GFS).

Regds,
SK

-------------------------------------------------------------------------
Take Surveys. Earn Cash. Influence the Future of IT
Join SourceForge.net's Techsay panel and you'll get the chance to share your
opinions on IT & business topics through brief surveys - and earn cash
http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV
_______________________________________________
Oscar-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/oscar-users

Reply via email to