high performance computing on graphics processing units: hgpu.org

This paper presents Model-T, an OS network stack designed to scale to terabit rates through pipelined execution of micro operations. Model-T parallelizes execution on multicore chips and enforces lockstep processing to maximize shared L2 data cache (d-cache) hitrate. Executing all operations without hitting main memory more than once (if at all) is the key design principle behind Model-T. We show a prototype implementation that indeed handles terabit rate network traffic when accessing only the L2 cache and processing only headers. Additionally, we present a more practical implementation of Model-T that is under development for Linux 2.6. Finally, we introduce an affordable test infrastructure based on general purpose graphics processor computation (GPGPU) that can replay network streams at PCI Express saturation rates (up to 128 Gbps), to benchmark Model-T and similar research network stacks.