Print

Print


  Hi Fons,

On Thu, Feb 03, 2005 at 08:01:57AM +0100, Fons Rademakers wrote:
>  very nice results indeed. Did each client go via the olbd or did they 
> contact directly the xrootd server?

  I expect that this was done by reading directly from the data server. Since
the redirector/olbd-network is used only to _find_ the actual data server from
which to read and then plays _no_ part in the actual reading itself, the
benchmark should remain unchanged. All you would get is the 5 second delay
at the beginning the first time the file is opened by one of the jobs on the 
system in any given 8 hour period.

  It isn't really a benchmark per se, but we know that the SLAC system (for
example) with the olbd scales up to 2000 clients easily, each typically 
reading from 1-2 files in any given moment.

> Also, although xrootd is primarily a data server, it would be nice to see a 
> write benchmark, like a few hundred clients trying to open write streams at 
> the same time to a bunch of xrootd servers.

  Andy certainly intended from the beginning that xrootd be usable for both
read and write. It was just BaBar's choice to deploy it primarily for the
large scale read-only systems as that was a higher priority issue. We had
other means of dealing with the write part of the problem. 

  Since late last fall we've been deploying a production load balanced
read/write system (with a lot of help from Remi, who has been using it). A 
quantitative benchmark would actually be quite useful at this point.

  Andy, using the BaBar applications for a write benchmark will probably be 
a bit limiting in that they are all very cpu-bound and hence won't necessarly 
give you a high write-rate/client. Probably the best thing is to use xrdcp
and copy a file from the local disk on the worker node into the system.

                                   Pete

> Andrew Hanushevsky wrote:
> >We finally got the xrootd performance benchmark together and put it on the
> >web:
> >
> >http://xrootd.slac.stanford.edu/presentations/Xrootd_Performance.htm
> >
> >Frankly, we were suprised (pleasantly so, I add). The results came out much
> >better that we expected.
> >
> >Andy



-------------------------------------------------------------------------
Peter Elmer     E-mail: [log in to unmask]      Phone: +41 (22) 767-4644
Address: CERN Division PPE, Bat. 32 2C-14, CH-1211 Geneva 23, Switzerland
-------------------------------------------------------------------------