Grant: Research Tools and Instruments - Category 1 (<$150,000)

A shared-memory manycore parallel computer is required to pursue research into parallel algorithms and architectures for diverse range of applications including FPGAs, GPUs, bioinformatics, data mining, verification and scheduling of complex systems. Ideally, the system would be at the leading-edge of modern mainstream systems, which today is 64 processors with 256GB of memory at roughly $85,000. In 3-5 years time, such a system is likely to be economically mainstream as a standard engineering workstation priced below $10,000. We plan to immediately utilize such a leading-edge platform to allow us to prepare in advance for what will be soon become economically mainstream. Performance scaling of traditional single-processor systems has slown dramatically since 2004 when processor clock frequency peaked around 3GHz. Instead, performance is now obtained through architectural innovation that exposes increased parallelism. As a result, mainstream processors now incorporate 2, 4 or even 8 cores in a single socket. In addition, hyperthreading allows 2 or more parallel program threads to be very finely interleaved on the same processor core - when one thread stalls waiting for slow main memory, one of the other threads immediately resumes and uses otherwise idle resources like the ALU or floating-point unit. It is even possible that a single core issues instructions from multiple threads in the same clock cycle. As a result, parallel processors have very quickly become mainstream. Very soon, we will have 16 to 64 cores available on these mainstream processors. To prepare for this wave of vastly increased parallelism, we need appropriate parallel algorithms and architectures. Parallel programming is more difficult and time-consuming than traditional sequential programming, a skill best learned in graduate school. Furthermore, increased effort is required to develop parallel algorithms which can utilize the massive levels of parallelism soon to be available. A shared-memory multiprocessor is required to enable this research into effective parallel algorithms and architectures.