Accelerating the Cloud with Heterogeneous Computing

Heterogeneous multiprocessors that combine multiple CPUs and GPUs on a single die are posed to become commonplace in the market. As seen recently from the high performance computing community, leveraging a GPU can yield performance increases of several orders of magnitude. We propose using GPU acceleration to greatly speed up cloud management tasks in VMMs. This is only becoming possible now that the GPU is moving on-chip, since the latency across the PCIe bus was too great to make fast, informed decisions about the state of a system at any given point. We explore various examples of cloud management tasks that can greatly benefit from GPU acceleration. We also tackle tough questions of how to manage this hardware in a multi-tenant system. Finally, we present a case study that explores a common cloud operation, memory deduplication, and show that GPU acceleration can improve the performance of its hashing component by a factor of over 80.