SQUAREMIX: A Faster Pseudorandom Number Generator for Dynamic-Multithreading Platforms

Many concurrency platforms offer a processor oblivious model of computation, where the scheduler dynamically distributes work across threads. While this is convenient, it introduces non\hyp determinism at runtime, which complicates debugging, since it precludes repeatability. Leiserson et al. [PPoPP '12] persuaded Intel to modify its C/C++ compiler, which provided the Cilk Plus concurrency platform, to include a feature called pedigrees, which enables determinism by uniquely identifying strands with low overhead. They used pedigrees to design a DPRNG called DOTMIX, which hashes a pedigree, then mixes the result into a random number for a given strand. Improving the efficiency of DOTMIX by using a faster hash function is an open problem put forth by Leiserson et al. [PPoPP '12]. We address this problem and improve the speed of the algorithm roughly by a factor of two, without sacrificing any statistical quality. Specifically, we replace the compression function used in DOTMIX with a faster universal hash function family due to Etzel et al. [CRYPTO '99] called Square Hash.