Applicative 2016 / Speakers

Update needs time: Scalable update to system data-structures

Procrastination has been a powerful technique to enable scalable data-structure access. RCU enables read-only data-structure operations to avoid explicit synchronization, thus to scale well with an ever-increasing number of cores. However, updates often require complicated synchronization between readers and updaters that does not scale well with more cores, and introduces undesirable latency spikes. During updates, memory reclamation is delayed until a future point when it is safe.

In this talk, we'll introduce a technique that uses local access to time as a global ordering on events to make updates cheap for preemptive systems. We'll discuss the multiple ways we've used time to synchronize between cores, thus minimizing shared memory synchronization.

About

Gabriel Parmer is an Associate Professor at The George Washington University in Washington DC. His research focuses on bringing both scalability, dependability, and security to embedded/IoT systems, and predictable latency guarantees to large-scale systems. Toward this, his group has researched and developed the Composite component-based OS for more than ten years. This research had produced many publications, multiple best-paper awards, and the NSF CAREER award.