Blue Gene/L
(BlueSky)

In 1999 IBM set out to develop a supercomputer that built on the basic architecture of the QCDOC machines but was capable of running a wider range of applications. The machine was to provide unprecedented computing performance in the Petaflops range and had a very low power consumption and cost.

The result was the first of the Blue Gene series of machines: Blue Gene L (L for Light). 

BlueSky, the BlueGene/L system procured by the University of Edinburgh, arrived in Edinburgh in December 2004 and was the first Blue Gene system available outside the USA. 

Although BlueSky consisted of only one rack (1024 compute nodes), it was an important machine for academic research. 

Lorna Smith was responsible for promoting Blue Gene use to academics within the University and for ensuring they were able to exploit the machine effectively. 

"The Blue Gene/L system provided local Edinburgh University staff with a significant local compute resource. The machine was a novel design and the large number of low clock speed processors created a challenge for some of our users. However with support many were able to access significant compute resource and achieve major advances in their science."

RESEARCH: INVESTIGATING THE GENETIC CAUSES OF BOWEL CANCER 

The aim of this collaborative project between EPCC and Colon Cancer Genetic Group (CCGG) of the MRC Human Genetics Unit at the Western General Hospital was to investigate genetic causes of bowel cancer. 

Understanding the relationship between genetic markers and colorectal cancer was thought to be a first step in identifying individuals at risk of the disease and to allow appropriate preventive measures to be taken. The studied dataset consisted of 565,000 genetic markers with real data from 1000 cancer cases and 1000 matched controls. 

On the Blue Gene system, the project was aimed at investigating the effect of each individual genetic marker, which required around 565,000 computations. 

During this phase EPCC successfully ported, parallelised and optimised the FORTRAN serial code on the BlueGene/L computer. The predicted serial run time for this O(N) problem was 10,800 days on a single CPU. The parallel code when run on 128 processors on BlueGene/L took 6.5 hours! The results led to a publication in the Nature Genetics.

EPCC's BlueSky was a single cabinet IBM eServer Blue Gene/L system containing 1024 compute nodes, consisting of dual-core 700MHz PowerPC 440 processors.

The low clock rate of the processors and their high flop/watt rate created a very power-efficient system while also minimizing the floor space required.

Each processor had a main memory of 512 MB of RAM and three levels of cache. The compute nodes could operate in two modes, either Co-processor (CO) or Virtual-Node (VN) mode. In the CO mode, one core dealt with computations while the other was used for communications. In the VN mode, both cores were used for computation and communications and as a result shared on-chip resources. Careful overlapping of communication and computation was required to achieve optimal performance.