|
CERN |
Two presentations were delivered
in the October meeting.
Anton Topurov
discussed the
Oracle Exadata storage system, which is a
rack-based concept with database servers, storage cells,
Infiniband and Ethernet switches. The storage system consists of
both hardware (from HP) and software (including Linux - from
Oracle). Such a system allows many Oracle I/O intensive
operations to be offloaded to the hardware, rather than being
done by the host database server. Anton explained that the
Oracle openlab team had been given remote access to a remote
system in the UK. The test time was limited, but the Swingbench
benchmark based on Table space creation for the PVSS application
indicated real improvements in execution time. More testing will
be done later, possibly also at CERN.
In the second presentation,
Gyorgy Balazs presented a
preliminary analysis of the performance monitoring data that
had been collected on about 60 LXbatch servers for almost 50
days. The data covered over 200000 production jobs and Gyorgy
related how he had used the logs of the batch system to identify
the groups responsible for each job. Since a dual-socket
quad-core server runs 8 or 9 different jobs at any moment in
time, it was often difficult to relate performance details to
the behaviour of a given group. In order to answer questions
like "is the group running 64-bit code?", Gyorgy selected
periods when at least 80% of the running jobs belonged to the
same experiment. At the end of his talk, there was a lively
discussion about how to best carry out such monitoring without
having to do too much guess work afterwards. A report will be
forthcoming, Gyorgy said in conclusion.
Sverre Jarp
CERN
|