Systems Investigations
Water-Cooled Racks
Blue Gene Application Performance Analysis and Benchmarking
Graphics Processor Computational Offloading
Cell Processor
Water-Cooled racks
As processors develop significant power needs and the density of processors per system becomes greater, difficulty arises in keeping machines cool through traditional air-cooled methods. The IT industry is witnessing the emergence of computer racks that rely on a liquid source, typically water or a refrigerant, to provide cool air to machines and to transport away the machines' exhaust heat. The GEaRS group has installed five water-cooled racks in the basement of the Computer Building to learn what problems and stumbling blocks could exist with this new technology, and how it can be deployed on a wider scale to address ITS-wide needs for a larger number of bigger servers.
Blue Gene Application Performance Analysis and Benchmarking
The IBM Blue Gene system is a revolutionary approach to deal with issues inherent in large-scale parallel computational systems, such as power consumption, heat load, and code scalability. To take advantage of the fact that power consumption increases in a super-linear manner to processor frequency, IBM designed Blue Gene to deliver more performance per watt through the use of many low-frequency processors and a small amount of RAM per compute node, rather than use fewer, high-frequency processors. To address inter-process communication between compute nodes, IBM implemented multiple co-existing network topologies in the system, which was optimized for different communication patterns. A potential disadvantage to this approach is that it could limit the types of problems able to be solved efficiently on the system, due to either limits of problem scalability or the amount of RAM a problem would need.
To investigate this, the GEaRS group identified Penn State researchers whose computational problems may benefit from using a Blue Gene machine. The GEaRS group also began constructive engagement with Argonne National Lab for access to their Blue Gene system and identified possible codes. Series of codes were tested from the life sciences, materials, chemistry, electrical and chemical engineering disciplines. Results of these tests were discussed in regular conference calls with the IBM Blue Gene team to obtain guidance for further optimizing applications, and to provide IBM with feedback on the benefits and limitations found in the Blue Gene system.
Graphics Processor Computational Offloading
Many scientific applications are very floating-point calculation intensive. Commodity processors, traditionally, have been designed as general use processors with only around 7% of their logic dedicated to floating-point operations. Graphics processors, however, dedicate up to 35% of their logic to floating-point operations and therefore can process many more floating-point calculations much more quickly. A modern graphics processor is capable of 360 gigaFLOPS (or GFLOPS - one billion floating-point operations per second), whereas a modern commodity processor is capable of only 18 GFLOPS. The GEaRS group is working with a leading-edge company that is developing software APIs to use graphics processors as offload engines for floating-point calculations. While these APIs are still in the beta stage, they offer the possibility of not only increasing the efficiency of solving existing problems, but also the possibility of enabling researchers to solve problems that are currently cannot be solved due to computational complexity.
Cell Processor
The Cell Processor is an exciting new processor from IBM designed to handle large amounts of floating point computation by utilizing an 9 core CPU design. One of the cores acts as the main core and the other 8 cores act as offload engines to the main core. It is estimated that the Cell Processor will reach around 256 GFLOPS. The GEaRS group has attended several seminars on the Cell Processor and has been working with IBM to gain access to early pre-production units.