Weaving Parallel Threads

As the speed of processors is starting to plateau, chip manufacturers are instead looking to multi-core architectures for increased performance. The ubiquity of multi-core hardware has made parallelism an important tool in writing performant programs. Unfortunately, parallel programming is still considered an advanced technique and most programs are written as sequential programs. We propose that we lift this burden from the programmer and allow the compiler to automatically determine which parts of a program can be executed in parallel. Historically, most attempts at auto-parallelism depended on static analysis alone. While static analysis is often able to find safe parallelism, it is difficult to determine worthwhile parallelism. This is known as the granularity problem. Our work shows that we can use static analysis in conjunction with search techniques by having the compiler execute the program and then alter the amount of parallelism based on execution speed. We do this by annotating the program with parallel annotations and using search to determine which annotations to enable. This allows the static analysis to find the safe parallelism and shift the burden of finding worthwhile parallelism to search. Our results show that by searching over the possible parallel settings we can achieve better performance than static analysis alone.