Eun-Jin Im and Katherine Yelick. Optimization of sparse matrix kernels for data mining. submitted to First SIAM Conf. on Data Mining, 2000.  Home/Search   Document Details and Download   Summary   Related Articles

....generated very machine specific optimizations like instruction selection can be left to the compiler. We propose to extend this approach to a much wider range of computational kernels by using compiler technology to automate the production of these SDCGs. Especially of interest are sparse kernels [34, 32]. Sparse matrix algorithms tend to run much more slowly than their dense matrix counterparts. For example, on a 250 MHz Ultrasparc II, a typical sparse matrix vector multiply implementation applied to a document retrieval matrix runs at less than 10 MFlops s, compared to 100 MFlops s for a ....

....and 400 MFlops s for matrix matrix. Major reasons for this performance difference include indirect access to the matrix and poor data locality in access to the source vector x in the sparse case. However, our current optimizations can speed sparse matrix vector multiplication up over five fold [34]. It is remarkable how the type of optimization depends on the matrix structure: For a document retrieval matrix, only by combining both cache blocking and multiplying multiple vectors simultaneously do we get a five fold speed; cache blocking alone yields a speedup of 2.5, and multiple vectors ....

Eun-Jin Im and Katherine Yelick. Optimization of sparse matrix kernels for data mining. submitted to First SIAM Conf. on Data Mining, 2000.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC