Online performance modelling of parallel applications

Summary:

Can we capture the performance characteristics of modern hardware in a simple model and use it in an operating system?

Motivation:

Trends in multicore hardware mean that in some performance characteristics modern computers are beginning to resemble the distributed message-passing machines from the early 1990s.

This project will revisit the performance models built for such systems (such as logP, logGP, etc.) and also more recent efforts (such as Roofline), with a view to establishing the usefulness of these models in resource allocation in a modern operating system (Barrelfish), designed for a diverse range of hardware platforms.

Goals:

1. Measure performance characteristics of parallel applications on modern hardware, and derive up-to-date performance models of such hardware.
2. Investigate whether such models can be used to predict the performance of benchmark applications on new hardware.
3. Build such hardware models online in the context of the Barrelfish operating system.
4. Explore how such models can be used to perform resource allocation in real time for parallel applications running on a multicore machine.

Technologies:

Performance modelling, C, parallel programming, OS design, scheduling.

Contact:

Timothy Roscoe