Incremental Parallel and Distributed Systems

摘要: Incremental computation strives for efficient successive runs of applications by reexecuting only those parts the that are affected a given input change instead recomputing everything from scratch. To realize benefits incremental computation, researchers and practitioners developing new systems where application programmer can provide an update mechanism changing data. Unfortunately, most existing solutions limiting because they not depart programming models, but also require programmers to devise (or dynamic algorithm) on per-application basis. In this thesis, we present parallel distributed enable real-world automatically benefit updates. Our approach neither requires departure current models programming, nor design implementation algorithms. achieve these goals, have designed built following systems: (i) Incoop — system MapReduce computation; (ii) Shredder GPU-accelerated storage; (iii) Slider stream processing platform slidingwindow analytics; (iv) iThreads threading library computation. experience with shows significant performance be achieved without requiring any additional effort programmers.