5th EGEE User Forum

gLite Porting to the Play Station 3 using ETICS for electronic High Throughput Screening (eHiTS)

12 Apr 2010, 17:12

3m

Aula (Uppsala University)

Poster Experiences from application porting and deployment Poster session

Speaker

Mr John Walsh (Trinity College Dublin)

Description

Since 2003 TCD has invested heavily in middleware porting, constantly engaging with the middlware development groups of the EU DataGrid (EDG), LHC computing Grid (LCG) and EGEE projects. In 2008, TCD ported gLite worker node to Yellow Dog Linux 6 on the Play Station 3, without data management. The node, built in ETICS, and tested in a production environment with gLite WMS job submissions were successfully submitted to it. The decrease in computational performance of porting the eHiTS software to the Grid is too expensive, so porting the Grid to other architectures such as the PS3 is important.

Conclusions and Future Work

An integrated build and test environment (including external projects
such as dCache, VDT and gridsite) on a very large set of platform types
will allow EGEE to produce sustainable gLite clients for worker nodes
and user interfaces. This is ongoing work at TCD, PSNC and CERN.

The exploitation and support of MPI on multi-core technology has proven
useful in recent years. However, the GPU and SPE are relatively
untouched at present, for highly parallelisable tasks in EGEE.

Detailed analysis

As a result of the accumulated expertise, TCD was invited in 2006 to
join a new EGEE-II integration and testing activity (SA3) as the main
portability partner for further development in this area, and then for
EGEE-III, Trinity College Dublin handled the porting and multi-platform
coordination.

It was found that applications running on hardware such as the Graphical
Processor Unit (GPU) and the Synergistic Processing Element (SPE) will
incur very large performance hits if ported to the Grid. In particular,
electric drug screening software such as eHiTS is one such application.

The meta-package generation facility of ETICS was proposed as a method
to quickly generate a minimal gLite worker node for the PS3 that would
accept glite WMS job submission but would not support data movement.
Producing such a solution first involved porting VDT globus to the PS3,
something not done before. The addition of a number of no architecture
RPMs and VOMS then allows the node to be contacted by a standard gLite
computational element (CE). However, a separate queue is required per
platform type to avoid confusion with other Linux platforms. A small
cluster of 7 nodes is now running production jobs.

Impact

The impetus to move the VDT globus source code stack into ETICS to allow
it to be ported to exotic platforms such as the PS3 and openSUSE has
resulted in the ability to build and maintain a version of VDT globus in
EGEE. This in itself is a useful achievement.

The more portable the EGEE Grid is, to platforms such as RedHat Linux,
Debian Linux, Mac OS X, Windows and PPC Linux (such as the PS3), the
more likely it is that users will use the Grid, since they will not have
to port their applications to the Grid.

eHiTS Lightning's computational speed is 26-fold to 60-fold faster on
the PowerPC Cell Broadband Engine (Cell B./E.) compared with the
equivalent application on an Intel-based processor. Therefore porting
the software to the Grid makes little sense from a purely economical
point of view.

Moreover, this immediately points to the fact that there is a whole
application space in existence for which the EGEE Grid could be used,
but isn't exploiting due to its lack of portability.

The use of GPU's is becoming much more cost effective. EGEE should
therefore pay close attention to the exploitation of such technologies
in the future.