ELITE: Near-Optimal Heuristics for Coflow Scheduling
Afaf Arfaoui, Rachid El-Azouzi, Francesco De Pellegrini, Cédric Richier, Jérémie Leguay · 2022 22nd IEEE International Symposium on Cluster, Cloud and Internet Computing (CCGrid) · 2022
Reducing Coflow Completion Time (CCT) has a significant impact on data-intensive application performance in datacenter networks. An efficient allocation of network resources allows for accelerating the computations to be performed. In this paper, we propose a new scheduler, named ELITE, to minimize the Weighted Coflow Completion Time (WCCT). Our scheduling algorithm is a 2-approximation of the optimal and the rate allocation can achieve a 4-approximation as long as the scheduling priority is respected. We also present a new rate allocation procedure, named RACO, that shows near-optimal performance when combined with our scheduling algorithms. We also propose a low complexity online scheduler, named LSPRT to attain near-optimal performance in online setting. With extensive simulations, we demonstrate the effectiveness of our algorithms by measuring the performance gain of ELITE and LSPRT over previous solutions in the literature. In particular, ELITE and LSPRT perform about 44 % better than Varys, while Sincronia achieves only 32 % against Varys.