Fast Estimation of Communication Cost for Thread Mapping in Computation Grids Using SVD
Ti-Hsin Wang, Tyng‐Yeu Liang, Chia-Hao Chang · 2008
Threads migration is performed in parallel programs in order to achieve communication minimization and efficient performance. Since the communication cost of the DSM system have to be measured through the calculation of type-sharing degree between threads, the cost of calculating such metric depends on the problem size of the application programs. The time complexity of calculating the communication costs caused by data sharing between m threads is O(m2n) while n is the number of pages shared by a pair of threads. This study proposed a fast estimation method based on the singular value decomposition (SVD) that could greatly reduce the amount of time required to calculate the type-sharing degree between thread pairs. Experimental results showed that this fast estimation algorithm provide a good solution toward a efficient way to improve the performance of thread migration on grid applications.