Mihir is a PhD candidate at the University of British Columbia, working with Andy Warfield and Bill Aiello. He is broadly interested in systems and has worked on multi-core scalability and performance, and on providing better security and performance isolation in virtualized environments. During his PhD, he also spent a couple of years at Coho Data, working on high-speed storage systems. Prior to graduate school, he worked in the security industry on detecting kernel-level malware and on black-box vulnerability detection for applications.
The datacenters behind popular cloud services are extremely resource-dense. A typical deployment has thousands of cores, terabytes of memory, gigabits of bandwidth, and petabytes of storage available per-rack. Datacenter economics require that providers share these resources across multiple customers for efficient utilization and as a means of providing price-competitive offerings. Shared infrastructure, however, risks cross-client interference and can result in degraded performance or data leaks, leading to outages and breaches. My work explores this tension with systems that provide security and performance isolation on shared hardware, while enabling efficient utilization and preserving the underlying performance of the devices.
This talk describes two such systems dealing with different resources: the first, Plastic, transparently mitigates poor scalability on multi-core systems caused by insufficient cache line isolation, which results in unnecessary memory contention and wasted compute cycles. Another one, Decibel, provides isolation in shared non-volatile storage and allows clients to remotely access high-speed devices at latencies comparable to local devices while guaranteeing throughput, even in the face of competing workloads.