Improving Performance of HPC Clouds via Hardware/Software Enclaves

Tuesday, March 31, 2020, 10:00 am - 11:00 am PDTiCal
Virtually via ZOOM
Networking & Cybersecurity Hosted Talk
Dr. Brian Kocoloski
Video Recording:



Cloud computing systems offer many benefits to applications, including the ability to provision virtually infinite resources in an on-demand fashion. This capability is increasingly attractive to the high performance computing (HPC) community, where workloads are increasingly "bursty" as well as data/event driven. However, while rapid innovation in cloud hardware has allowed providers to offer compelling high performance cloud instances that could in theory support these workloads, issues remain that limit performance and hinder wider spread adoption of cloud resources in HPC communities. This talk first focuses on one such issue: the system software stacks commonly used in clouds introduce performance variation. Specifically, we show that commodity operating system kernels introduce variation as workload contention rises. In contrast to previous works, however, we demonstrate that this variation arises due to contention for system software resources (e.g., locks, page caches, memory allocators) as opposed to contention for hardware. Our work introduces the notion of the "kernel surface area," a measurement of how many resources the kernel is managing, and shows that software introduces more variation as the kernel surface area increases. An additional finding of our research is that OS containers, due to their reliance on a shared kernel, are more susceptible to this issue than virtual machines.

We will then present the design and implementation of a new multi-kernel system, Hobbes, that is designed to limit this variation. We show how Hobbes leverages hardware resource partitioning mechanisms provided in modern hardware to provision "enclaves." Enclaves allow applications to execute in a completely isolated environment; workloads executing in two different enclaves share neither hardware nor system software. We demonstrate how we can also leverage hardware virtualization to create feature-rich enclaves whose capabilities outstrip those provided in other multi-kernel like systems.

Finally, we conclude with a discussion of the additional ways in which we are using hardware partitioning mechanisms to improve cost-efficiency for HPC/cloud applications.

Speaker Bio:

Brian Kocoloski is an assistant professor in the Computer Science and Engineering department at Washington University in St. Louis. His research interests include performance and cost efficiency for high performance computing (HPC) applications, particularly in the context of shared cloud-based systems. He received his Ph.D. in Computer Science from the University of Pittsburgh in 2018.

ISI Host:

Terry Benzel, Division Director - Networking & Cybersecurity


« Return to Upcoming Events