BARR: Congestion aware scheduling algorithm for Network-on-Chip router

Abstract

Scheduling algorithm is crucial to the performance of the Network-on-chip router. Different from traditional scheduling algorithms that concentrate on local fairness, we propose a congestion-aware scheduling algorithm based on input buffer of downstream router. The scheduling algorithm keeps a match dynamically between input and output by detecting the flits number to be transferred in the same packet. It can reduce network congestion especially under heavy traffic loads. Compared to RRM and iSLIP algorithm, the new scheduling algorithm can increase the saturation throughput by 8.2% and reduce the average communication latency by 7.8% under non-uniform traffic.

Flit Scheduling for Cut-Through Switching: Towards Near-Zero End-to-End Latency

“…The paper, however, assumes that every flow has its own priority class (instead of being mapped to one of finite priority classes), which is unrealistic. The buffer-based adaptive round-robin scheduling algorithm presented in [16] utilizes the buffer state of downstream routers to reduce the chance of congestion. Such studies in NoC, however, assumed not only a small buffer size at each router but also a well-defined topology, mainly 2D-mesh, which are only valid in the NoC environment.…”

“…[25] proposed iSLIP, an iterative and round-robin scheduling algorithm for ATM switches aiming at avoiding headof-line blocking and starvation at input queues. Although originally designed for ATM, iSLIP has been also widely used for DCN (e.g., Cisco Nexus 5000 series [26]) and for NoC [16], [27]. In Section V, we will use iSLIP as one of reference algorithms to compare with our proposed algorithms, not only due to its popularity (in ATM, DCN, and NOC), but also due to its generality to fit into a random topology like ATM.…”

“…Next, f EDF-hop deals with the perhop deadline of packet P i,k at an OQ, which is defined as the to-destination deadline divided by l i,k path , to capture the impact of the number of the remaining links to the destination. Then at an OQ, the f EDF (either f EDF-dst or f EDF-hop) algorithm calculates the deadline of every existing packet, 16 and the queued flit belonging to the packet with the minimal value (i.e., the earliest deadline) is chosen to transmit with the transmission rate equal to the output link rate. As a result, the computational complexity of f EDF is given as O (n) + O (n) = O (n), which includes calculating the deadlines of n queued packets and comparing them to find the smallest deadline.…”

“…A set of jobs is schedulable if there exist one or more solutions to schedule all the jobs within their deadlines 16. An existing packet implies the one whose head flit has arrived earlier but tail flit has not yet been transmitted to the next router.…”