Multicore Algorithm Design
Roman Pearce, CECM, Simon Fraser.Wednesday February 10th, in K9509 at 3:30pm.
Abstract. In this talk we will describe the principles at work in the design and implementation of high performance, fine grained parallel programs for multicore processors. In addition to example code that you can try yourself, we will discuss two real problems where these techniques were successfully used. Given two sparse polynomials f = f_1 + f_2 + ... + f_n and g = g_1 + g_2 + ... + g_m, our first problem is to compute the product f x g. We compute, sort, and merge all pairwise products f_i x g_j in parallel while writing out the result. Our next problem is to divide f/g and compute the quotient q with f - q g = 0. The quotient is computed term by term while it is being multiplied by g and subtracted from f in parallel, with no redundant work. Both algorithms achieve superlinear speedup on an Intel Core i7 CPU with 4 cores.