TCU: A Multi-Objective Hardware Thread Mapping Unit for HPC Clusters

Meeting multiple, partially orthogonal optimization targets during thread scheduling on HPC and manycore platforms simultaneously, like maximizing CPU performance, meeting deadlines of time critical tasks, minimizing power and securing thermal resilience, is a major challenge because of associated scalability and thread management overhead. We tackle these challenges by introducing the Thread Control Unit (TCU), a configurable, low-latency, low-overhead hardware thread mapper in compute nodes of an HPC cluster. The TCU takes various sensor information into account and can map threads to 4-16 CPUs of a compute node within a small and bounded number of clock cycles in round-robin, single- or multi-objective manner. The TCU design can consider not just load balancing or performance criteria but also physical constraints like temperature limits, power budgets and reliability aspects. Evaluations of different mapping policies show that multi-objective thread mapping provides about 10 to 40% less mapping latency for periodic workloads compared to single-objective or round-robin policies. For bursty workloads under high load conditions, a 20% reduction is achieved. The TCU macro has a mere 9% hardware area overhead and achieves more than 150k thread mappings per second on an FPGA prototype of a RISC quad-core compute node operating at moderate 50 MHz. A 45 nm technology ASIC realization of TCU can operate well above 1 GHz and support up to 3.15 million thread mappings per second.     «

Meeting multiple, partially orthogonal optimization targets during thread scheduling on HPC and manycore platforms simultaneously, like maximizing CPU performance, meeting deadlines of time critical tasks, minimizing power and securing thermal resilience, is a major challenge because of associated scalability and thread management overhead. We tackle these challenges by introducing the Thread Control Unit (TCU), a configurable, low-latency, low-overhead hardware thread mapper in compute nodes o...     »