The IBM Blue Gene Supercomputing Project

About the Project

The IBM Blue Gene supercomputing project is a broad effort by IBM to construct made-to-order supercomputers from a fabric of generally off-the-shelf parts. The systems under design range in scale from beta versions of BlueGene/L, which is a general purpose cluster intended to compute in the range of 360 trillion floating point operations per second (TeraFLOPS), up to BlueGene/Q, which will eventually scale to 3000 TeraFLOPS.

The "gene" portion of the Blue Gene name derives from the intended use of the Blue Gene clusters: Computational biology, specifically, protein folding. This highly-parallelizable but computationally expensive problem is seen as a cutting-edge field in genetics and drug research, and large quantities of research monies are available for computational resources, motivating the build out.

Current and Planned Installations

As of this writing, only BlueGene/L has significant build out, though BlueGene/C is expected to be completed soon. BlueGenes/{P,Q} are still tentative, and are likely to be larger-scale implementations of BlueGene/L in the opinion of the node author.

Why the confusion about the name?

Many people mean BlueGene/L when they say "Blue Gene." While BlueGene/L is the only current functional Blue Gene computer, it is not the only architecture represented under the Blue Gene umbrella, nor is it the only implementation that will exist.

Architectural Considerations

The primary design considerations in the Blue Gene clusters are very different from previous efforts, and this is strongly reflected in almost every aspect of the implementation of BlueGene/L. In the past, the focus of supercomputing clusters has been on performance above all else, but it has become unreasonable to expect non-governmental entities to foot the enormous power and space costs incurred in something like the NEC Earth Simulator, the previous "world's fastest computer" before BlueGene/L knocked it off the top spot of the Top500 supercomputing list. The Earth Simulator has the following statistics.

  • Space: Occupies its own facility (3200 sq.m.)
  • Power: About 10 Megawatts
  • Performance: 40 TeraFLOPS (4 GigaFLOPS/sec/Kilowatt)

By contrast, the half-size (32 k-node) BlueGene/L "DD2 Beta" at LLNL has the following.

  • Space: Occupies a single room (~130 sq.m., approx 37 standard racks)
  • Power: About 0.6 Megawatts
  • Performance: 70.7 TeraFLOPS (118 GigaFLOPS/sec/Kilowatt)

As can be seen, BlueGene/L is far more power and space efficient per FLOPS than the Earth Simulator. While some of this can be explained by the typical transistor scaling that goes into all speed improvements, the rate of improvement far outstrips the technological basis. Instead, the improvements come from the architectural choices made in BlueGene/L's design, namely in its use of low power embedded cores that operate at a low clock rate. By using PowerPC 440 processors instead of larger, more full-featured processors, the BlueGene/L designers were able to shove 1024 dual-core chips into a single cabinet (rack), along with all of the necessary infrastructure to connect that cabinet to the network. This functionally makes BlueGene/L an enormous blade server, with all of the low-power and high-density considerations that flow from that design decision. The speed loss per processor is compensated for by using 65536 of the dual-core chips in the system. Further discussion of this is (or will be) available at the BlueGene/L node, as the minutae are too detailed for inclusion here. The primary point of note in the discussion is that the Blue Gene project emphasizes the use of large numbers of simple processors over small numbers of complex processors, saving power as it does so.

References

  • My own experience as a graduate researcher at Penn State, working in the low-power supercomputing research group in Computer Science and Engineering.
  • The "Blue Gene" Wikipedia article, for information on planned build sites.
  • The LLNL presentation on BlueGene/L at http://www.llnl.gov/asci/platforms/bluegenel/images/BGL.pdf
  • Various LLNL/IBM publications on the implementation and aforementioned minutae of BlueGene/L.

Class adjourned.

Log in or register to write something here or to contact authors.