A hybrid multicast deadlock-free scheme for virtualization at the NoC level

Abstract

A new hybrid multicast deadlock-free scheme is proposed to enhance the multicast capability. In the proposed scheme, the small-sized multicast packet is routed in a deadlock-free way by packet-buffered and asynchronous replication, and the large-sized multicast packet is transferred under the control of centralized allocator, which restricts the number of concurrent multicast with large-sized packet transmission. The virtualization at the Network-on-Chip (NoC) level is also taken into consideration, that the allocator reserves respective counters for each sub-network. According to result of the experiment under the real workload traces, the performance reduction caused by centralized allocator is negligible. The router and the allocator are synthesized in Chartered 90nm CMOS technology. Compared with the only packet-buffered scheme, the allocator only consumes extra 0.079% to 0.2972% area overhead according to the different setting (routing mechanism, network size, number of supported sub-network) while offering the ability of large-sized multicast packet transmission.