Massive Main Memory for the Masses

Trends in DRAM scaling and new kernel-bypass-based networking have given rise to fast disaggregated storage that can process millions of requests per second per server with just a few microseconds of delay. These systems are fast in large part because they are dumb. Their stripped down data models make them easy to implement and optimize, but several things still limit the practicality of these systems. Simple data models force applications to "pay back" much of the performance they gain, since applications have to make many round-trips to perform simple operations and since they must perform all processing client-side. Furthermore, these systems are expensive; typically, smaller scale applications can't afford their own deployments of in-memory storage.  In this talk, I will cover ongoing work from my lab that tackles these weakness to make low-latency in-memory storage more practical. Our efforts focus on resource provisioning under massive multi-tenancy to make these types of systems amenable to cloud-scale consolidation, and programmability/extensibility that preserves performance gains by minimizing isolation costs for tenant-provided in-storage functions.

Ryan Stutsman is an Assistant Professor at the University of Utah.  He received his Ph.D. from Stanford in 2013 working on the RAMCloud low-latency in-memory data center storage system. After that, he was a Postdoctoral Researcher in the Databases group at Microsoft Research where he worked on Deuteronomy, a project to support highly scalable cloud storage transactions.  Ryan's current work is funded by a recent NSF CAREER award and targets new inter-tenant isolation techniques for fast, kernel-bypass-based storage systems.