Intel announces accelerator card

Today at 12:00 at the International Supercomputing Conference ISC2010 here in Hamburg, Intel announced a new accelerator card featuring 32 cores and up to 2 GB RAM which can run 128 threads via hyperthreading. Looks like the accelerator wars have begun. They showed a LU decomposition test running on the card and claimed it to be the fastest decomposition at the moment with more than 500 GFlop/s. Pretty impressive. I was attending the keynote from Kirk B. Skaugen
Vice President, Intel Architecture Group & General Manager, Data Center Group, Intel, USA who demoed the accelerator card and presented a complete software stack including C and FORTRAN compilers.

R on the SGI Altix 4700

Here at LRZ we have a large SGI Altix 4700 Supercomputer. Lately we have compiled an MPI-enabled version of R for this beast. The machine sports 9728 cores connected with a ccNUMA link and maxes out at 62 TFlop/s. With R we were able to use a whooping 4000 cores on the machine and still get a decent performance.

MPI programming with R is a breeze. You can use the foreach packages from Revolutions and just transfer your serial code to the supercomputer, replace all %do% with %dopar% and submit your jobs and you are done.

Here is a short tutorial how to start programming on supercomputers with R.

Monads and BSP

Today there was a post on reddit about a nice syntax for monads in python and it makes me wonder whether monads coudl help a lot in BSP. What they essentially provide is an abstraction for map and filter operations on arbitrary data sets. This leads me to this question: "Can the BSP model be rewritten in terms of monads?"

BSP is map-reduce?

Bulk synchronous parallel programming is a programming model that divides the execution flow into two alternating phases:

1. local computation
2. global communication and synchronisation
   

While MPI itself is rather hairy to program, BSP facilitates parallel programming a lot. It automatically avoids deadlocks as all processes wait for each other after the communication step. In that respect it looks very much like the map-reduce model that google uses, but seems to be a generalisation as it also allows both global and distributed arrays. Also a BSP program is a program for N processors while map-reduce is a model for N-1 slave processors and one additional master. In the weeks to come I will tell more about the experience I will make with that programming model. Stand by.

Python and MPI

SGI delivers the machine with preinstalled linux with kernel 2.6 which of course includes a python interpreter. It is version 2.4.2. Not bleeding edge but pretty up to date. But of course no support of MPI or PVM. So I dabbled in installing some MPI module. Currently there are a handful of MPI modulaes available but most of them seem defunct or dead. Finally I found a simple MPI module in the Scientific Python module by Konrad Hinsen (http://dirac.cnrs-orleans.fr/plone/software/scientificpython) which installed itself flawlessly and contains a nice module called BSP aka bulk synchronous parallel programming.

Booting up

Started my new job at the Leibniz Supercomputing Centre in Munich. One of my goals will be the implemetation of a good scripting environment on the HLRB2. It is an SGI Altix system with 9728 cores. Looks kinda cool... see:
http://www.lrz-muenchen.de/services/compute/hlrb/hardware/hardware.html