UC San Diego
Simulation to scale of the HELIOS system
- Author(s): Subramanya, Vikram
- et al.
Within the space of a few years, IT companies with need for large-scale datacenters are considering alternatives to building traditional central datacenter facility. They have embraced a modular datacenter (also called pod, a smaller mobile datacenter) as the building block for reasons such as efficient cooling, power savings, mobility, easier deployment and administration. However, delivering scalable inter-pod bisection bandwidth remains a challenge. Current solutions that promise full bisection bandwidth between arbitrary pair of hosts can be expensive and power-intensive for interconnecting modular datacenters. Our recent work [FPR+10] proposed Helios, a scalable switch architecture for interconnecting modular datacenters using a hybrid core switch array consisting of electrical and optical switches. We show that Helios delivers a significant reduction in cost, cabling complexity, and power while providing scalable on-demand bandwidth to the communicating pair of hosts. In this thesis, we propose a TCP flow-based simulator for the Helios architecture. We describe the design/implementation, and validate the correct behavior of the simulator with the testbed execution of Helios prototype. We then evaluate the performance of Helios system against a variety of parameters - like scale (number of pods), link aggregation, and the composition of the core switch array - for differing communication patterns. The primary goal of this tool is to provide insights into how Helios might scale. Possible applications of this tool include network planning, searching newer optimizations, identifying bad application performance, planning VM migration such that the VMs better utilize Helios infrastructure