Strict and Relaxed Sieving for Multi-Core Programming

In Codeplay’s Sieve C++, the programmer can place code inside a “sieve block” thereby instructing the compiler to delay writes to global memory and apply them in order on exit from the block. The semantics of sieve blocks makes code more amenable to automatic parallelisation. However, strictly queueing writes until the end of a sieve block incurs overheads and is typically unnecessary. If the programmer can assert that code within a sieve block does not write to and then subsequently read from a global memory location, the sieve semantics can be relaxed: writes can be executed at any point from their originally scheduled time until the end of the block. We present experimental results demonstrating the benefits of relaxed sieving on an x86 multi-core system.