Unsynchronized Techniques for Approximate Parallel Computing

We present techniques for obtaining acceptable unsynchronized parallel computations. Even though these techniques may generate interactions (such as data races) that the current rigid value system condemns as incorrect, they are engineered to 1) preserve key data structure consistency constraints while 2) producing a result that is accurate enough, often enough. And because they contain no synchronization, they eliminate the synchronization overhead, parallelism reduction, and failure propagation typically associated with synchronization mechanisms. We also identify a class of computations that interact well with our techniques. These computations combine multiple contributions to obtain a composite result. Redundancy in the contributions enables the unsynchronized computation to produce an acceptably accurate result even though it may, in effect, drop contributions. Thus our unsynchronized techniques (either directly or indirectly) are often acceptably accurate for the same reason as other successful approximate computing techniques such as task skipping and loop perforation.