Defense Event

Energy-Efficient Acceleration of Asynchronous Programs

Gaurev Chadha

Monday, December 14, 2015
3:00pm - 5:00pm
3725 Beyster Bldg.

Add to Google Calendar

About the Event

Asynchronous or event-driven programming has become the dominant programming model in the last few years. In this model, computations are posted as events to an event queue from where they get processed asynchronously by the application. A huge fraction of computing systems built today use asynchronous programming. All the Web 2.0 JavaScript applications (e.g., Gmail, Facebook) use asynchronous programming. There are now more than two million mobile applications available between the Apple App Store and Google Play, which are all written using asynchronous programming. Distributed servers (e.g., Twitter, LinkedIn, PayPal) built using actor-based languages (e.g., Scala) and platforms such as node.js rely on asynchronous events for scalable communication. Internet-of-Things (IoT), embedded systems, sensor networks, desktop GUI applications, etc., all rely on the asynchronous programming model. Despite the ubiquity of asynchronous programs, their unique execution characteristics have been largely ignored by conventional processor architectures, which have remained heavily optimized for synchronous programs. Asynchronous programs are characterized by short events executing varied tasks. This results in a large instruction footprint with little cache locality, severely degrading cache performance. Also, event execution has few repeatable patterns causing poor branch prediction. This thesis proposes novel processor optimizations exploiting the unique execution characteristics of asynchronous programs for performance optimization and energy-efficiency. These optimizations are designed to make the underlying hardware aware of discrete events and thereafter, exploit the latent Event-Level Parallelism present in these applications. Through speculative pre-execution of future events, cache addresses and branch outcomes are recorded and later used for improving cache and branch predictor performance. A hardware instruction prefetcher specialized for asynchronous programs is also proposed as a comparative design direction.

Additional Information

Sponsor(s): Satish Narayanasamy and Scott Mahlke

Open to: Public