03 并行程序设计基础.ppt

* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 14 * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * Parallel Processing Evaluation .8 Speedup measurement formula: Elapsed time of a uniprocessor Speedup = Elapsed time of the multiprocessors * Parallel Processing Evaluation .9 For example, if the uniprocessor took 40 seconds to perform a task, and two parallel systems took 20 seconds, then the value of speedup = 40 / 20 = 2. Value 2 indicates the ideal of linear speedup: when twice as much, hardware can perform the same task in half the time. * Parallel Processing Evaluation .10 Figure. Linear and actual speedup * Parallel Processing Evaluation .11 Amdahl’s Law Amdahls Law is a law governing the speedup of using parallel processors on a problem, versus using only one sequential processor. Amdahl’s law attempts to give a maximum bound for speedup from the nature of the algorithm: Amdahl 定律 出发点： 固定不变的计算负载； 固定的计算负载分布在多个处理器上的， 增加处理器加快执行速度，从而达到加速的目的。 * Parallel Processing Evaluation .12 Amdahl’s Law S: purely sequential part P: parallel part S + P = 1 (for simplicity) Maximum speedup = S + P S + P n = 1 * * Parallel Processing Evaluation .13 Gustafson’s Law If the size of a problem is scaled up as the number of processors increases, speedup very close to the ideal speedup is possible. That is, a problem size is virtually never independent of the number of processors. * Parallel Processing Evaluation .14 Gustafson’s Law Maximum speedup = S + ( P * n ) S + P = n + (1 - n) * S Gustafson 定律 出发点： 对于很多大型计算，精度要求很高，即在此类应用中精度是个关键因素，而计算时间是固定不变的。此时为了提高精度，必须加大计算量，相应地亦必须增多处理器数才能维持时间不变； 除非学术研究，在实际应用中没有必要固定工作负载而计算程序运行在不同数目的处理器上，增多处理器必须相应地增大问题规模才有实际意义。 * Parallel Processing Evaluation .15 Exam