Resilient and efficient communication in many-core systems using network coding

Due to technology scaling, the number of processor cores on a chip constantly increases. Already today, we reach the domain of so called many-core systems-on-chip. However, this advance comes at the cost of reliability, which especially affects the communication performance of the underlying network-on-chip. Today´s resiliency concepts for network-on-chip like automatic repeat request with retransmission are not feasible anymore and lead to long latencies and high network load. In this paper, we propose an on-chip transmission concept based on random linear network coding to provide high resiliency and an efficient communication in many-core processors at the same time. The concept offers a flexible and efficient computable coding scheme, which is well suited for on-chip communication and allows to exploit the path diversity of large networks. First, we use a flit-level cycle-accurate simulation model to investigate the performance potential of the proposed transmission scheme on a network-on-chip with 64 cores. Second, we propose an analytic model for random linear network coding in network-on-chip with retransmission, which is able to provide a very accurate performance estimation close to the cycle-accurate simulation. Finally, we apply the analytic model to investigate the performance potential on the large-scale, assuming a processor with 1024 cores.