Revisiting the cache miss analysis of multithreaded algorithms
LATIN'12 Proceedings of the 10th Latin American international conference on Theoretical Informatics
| Hi-index | 0.00 |
We consider algorithms for a multicore environment in which each core has its own private cache and false sharing can occur. False sharing happens when two or more processors access the same block (i.e., cache-line) in parallel, and at least one processor writes into a location in the block. False sharing causes different processors to have inconsistent views of the data in the block, and many of the methods currently used to resolve these inconsistencies can cause large delays. We analyze the cost of false sharing both for variables stored on the execution stacks of the parallel tasks and for output variables. Our main technical contribution is to establish a low cost for this overhead for the class of multithreaded block-resilient HBP (Hierarchical Balanced Parallel) computations. Using this and other techniques, we develop block-resilient HBP algorithms with low false sharing costs for several fundamental problems including scans, matrix multiplication, FFT, sorting, and hybrid block-resilient HBP algorithms for list ranking and graph connected components. Most of these algorithms are derived from known multicore algorithms, but are further refined to achieve a low false sharing overhead. Our algorithms make no mention of machine parameters, and our analysis of the false sharing overhead is mostly in terms of the the number of tasks generated in parallel during the computation, and thus applies to a variety of schedulers.