Abstract

The “free lunch” of exponentially increasing sequential software performance is said to be over: all major processor manufacturers are chasing increasing degrees of hardware parallelism instead of ever-higher clock speeds. We need parallel software to make good use of these processors, but parallel programming has a notoriously steep learning curve. What can the computer science research community do to make things easier?

One tempting approach is to automatically build parallel implementations from sequential programs. This led to a lot of work in the 1980s in the functional programming community, and the exploration of data-flow processor architectures that could exploit the very fine grained parallelism that was achieved. In this talk I’ll describe some work we’ve been doing returning to this question and investigating whether we can achieve automatic parallelisation of software when using today’s multi-core processors.

Our approach is to (i) profile the execution of a program, (ii) from this to identify pieces of work which are promising sources of parallelism, (iii) use spare processor cores to speculatively perform this work, (iv) detect at run-time any attempt to perform operations that would reveal the presence of speculation. Our results show 10-80% speed-ups on several programs. This is clearly not a replacement for careful manual parallelisation, but may provide a way of squeezing out a little extra performance without much effort: free coffee, if not a free lunch.