Hybrid and optical switching scheduling algorithms in data center networks

Abstract

As a cost-effective approach to the data center network scalability problem, hybrid-switched data center networks have received considerable research attention recently. A hybrid-switched data center network employs a much faster circuit switch that is reconfigurable with a nontrivial cost, and a much slower packet switch, to interconnect its racks of servers. The following optimization problem is the focus of most of the research works on hybrid switching: Given a traffic demand (between the racks), how to properly schedule the circuit switch so that it removes most of the traffic demand, leaving little for the slower packet switch to handle. The objective of this dissertation research is to design high-performance low-complexity scheduling algorithms for optical switches in hybrid-switched data center networks to boost the throughput performance.

In this thesis, we propose three algorithms for single optical switch scheduling that exploit different methodologies. We analyze, simulate, and summarize their respective merits on different metrics, such as the throughput performance, the computational complexity, the applicable conditions, etc. This thesis also investigate another closely related research problem about the optical and hybrid switching scheduling, that is, when the racks of servers are connected by multiple independent (i.e., parallel) optical switches, how to split the overall traffic demand into sub-workload and give them to the parallel optical switches as their respective workloads. We formulate it as a matrix split and balance problem and develop a general algorithm (not limited to optical switch scheduling) to split a matrix into balanced and sparse matrices. Our evaluation results show that, using this matrix split algorithm, parallel optical switches deliver balanced and ideal throughput performance under various system parameter settings and various traffic demands.