Communication avoiding parallel algorithms for amorphous problems

Amorphous;Symbolic;Parallel;Communication Avoiding;Linear Algebra;Tropical Semi-Ring;Single-Source Shortest Path;Dynamic Programming;Longest Common Subsequence;Smith-Waterman;Needleman-Wunsch;Viterbi Decoder;Parallel Notations

Maleki, Saeed

关键词: Amorphous;    Symbolic;    Parallel;    Communication Avoiding;    Linear Algebra;    Tropical Semi-Ring;    Single-Source Shortest Path;    Dynamic Programming;    Longest Common Subsequence;    Smith-Waterman;    Needleman-Wunsch;    Viterbi Decoder;    Parallel Notations;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/88981/MALEKI-DISSERTATION-2015.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Parallelizing large sized problem in parallel systems has always been a challenge for programmer. This difficulty is caused by the complexity of the existing systems as well as thetarget problems. This is becoming a greater issue as the data sizes are constantly growing and as a result, larger parallel systems are required.Graph algorithms, machine learning problems and bio-informatics methods are among the many ever-growing problems. These group of problems are amorphous, meaning thatmemory accesses are unpredictable and the application usually has a poor locality. Therefore, synchronizations in these problems are specially costly since all-to-all communications arerequired and delivering an efficient parallel algorithm becomes more challenging. Another difficulty with these problems is that the amount of parallelism in them is limited whichnaturally makes them hard to parallelize. This is due to complicated data-dependences among the data elements in the algorithm.Writing parallel algorithms for these problems, on the other hand, are specially difficult since an amorphous problem can be expressed in several dramatically different ways. Thisis because of complex data dependences which are statically unknown and therefore, manyunique parallel approaches exist for a single problem. Consequently, programming each single approach requires starting from scratch which is time consuming.This thesis introduces several ways to avoid costly communications in amorphous problems by compromising from the computation. This means that we can increase the total amount of work done by the processors to avoid synchronizations in an algorithm. Thisis specially effective in large clusters since there is a massive computing power with very costly communications. These approaches, clearly, have a trade off between computation and communication and in this thesis, we study these trade offs as well. Also, we propose a new language to express the proposed algorithms to overcome the programming difficulty ofthe problems by providing tunable parameters for performance.