Using FPGAs to Simulate Novel Datacenter Network Architectures At Scale

The tremendous success of Internet services has led to the rapid growth of Warehouse-ScaleComputers (WSCs). The networking infrastructure has become one of the most vital componentsin a datacenter. With the rapid evolving set of workloads and software, evaluatingnetwork designs really requires simulating a computer system with three key features: scale,performance, and accuracy. To avoid the high capital cost of hardware prototyping, manydesigns have only been evaluated with a very small testbed built with off-the-shelf devices,often running unrealistic microbenchmarks or traces collected from an old cluster. Manyevaluations assume the workload is static and that computations are only loosely coupledwith the very adaptive networking stack. We argue the research community is facing ahardware-software co-evaluation crisis.In this dissertation, we propose a novel cost-efficient evaluation methodology, calledDatacenter-in-a-Box at Low cost (DIABLO), which uses Field-Programmable Gate Arrays(FPGAs) and treats datacenters as whole computers with tightly integrated hardware andsoftware. Instead of prototyping everything in FPGAs, we build realistic reconfigurable abstractedperformance models at scales of O(10,000) servers. Our server model runs the fullLinux operating system and open-source datacenter software stack, including productionsoftware such as memcached. It achieves two orders of magnitude simulation speedup oversoftware-based simulators. This speedup enables us to run the full datacenter software stackfor O(100) seconds of simulated time. We have built a DIABLO prototype of a 2,000-nodesimulated cluster with runtime-configurable 10 Gbps interconnect using 6 multi-FPGA BEE3boards