A locality-aware shuffle optimization on fat-tree data centers

Shuffle is a data exchanging phase that is always inserted between two adjacent computations to deliverintermediate results in data centers. It generates a burst of traffic that exhausts the network bandwidth, debasing the availability of the core layer facilities in fat-tree topologies. Previous researches follow either flow inhibition or infrastructural upgrading to achieve a high utilization of core network resources. However, dynamic pressure from shuffle burst introduces more unpredictable usage of core network that disturbs the global locality-based optimization on task schedule. In this work, we reduce the core bandwidth consumption by scheduling the location of adjacent computing workers based on our proposed distance model that uses a similarity-based distance to evaluate the dynamic distance between fat-tree leaf-nodes. Task assignment further utilizes this distance to schedule workers to avoid high intensity usage of core network resources. This design improves the performance of shuffle phase in popular on-data-center algorithms as well as maintains infrastructural inexpensiveness of their fat-tree topology. The proposed models are evaluated on a semi-physical simulation test platform and compared to state-of-the-art solutions, such as Space Shuffle and Scalable Shuffle. The results show that our design achieves an up to 18% speedup on shuffle procedure and a 23% extension of network capacity. In addition, a significant mitigation of congestion can be obtained on the bottleneck of core network.