Showing papers by "Rezaul Chowdhury published in 2004"

Cache-oblivious shortest paths in graphs using buffer heap

Abstract: We present the Buffer Heap (BH), a cache-oblivious priority queue that supports Delete-Min, Delete, and Decrease-Key operations in O(1overB log2NoverB) amortized block transfers from external memory, where B is the (unknown) block-size and N is the maximum number of elements in the queue. As is common in cache-oblivious algorithms, we assume a 'tall cache' (i.e., M = Ω(B1 + e), where M is the size of the main memory). We also assume the Decrease-Key operation only verifies that the element does not exist in the priority queue with a smaller key value, hence it also supports the insert operation in the same amortized bound. The amortized time bound for each operation is O(log N). We also present a Cache-Oblivious Tournament Tree (COTT), which is simpler than the Buffer Heap, but has weaker bounds.Using the Buffer Heap we present cache-oblivious algorithms for undirected and directed single-source shortest path (SSSP) problems for graphs with non-negative edge-weights. On a graph with V vertices and E edges, our algorithm for the undirected case performs O(V + EoverB log2VoverB) block transfers and for the directed case performs O((V + EoverB) . log2VoverB) block transfers. The running time of both algorithms is O((V + E). log V).For both priority queues with Decrease-Key operation, and for shortest path problems on general graphs, our results appear to give the first non-trivial cache-oblivious bounds.

The Limits of Alias Analysis for Scalar Optimizations

Abstract: In theory, increasing alias analysis precision should improve compiler optimizations on C programs. This paper compares alias analysis algorithms on scalar optimizations, including an analysis that assumes no aliases, to establish a very loose upper bound on optimization opportunities. We then measure optimization opportunities on thirty-six C programs. In practice, the optimizations are rarely inhibited due to the precision of the alias analyses. Previous work finds similarly that the increased precision of specific alias algorithms provide little benefit for scalar optimizations, and that simple static alias algorithms find almost all dynamically determined aliases. This paper, however, is the first to provide a static methodology that indicates that additional precision is unlikely to yield improvements for a set of optimizations. For clients with higher alias accuracy demands, this methodology can help pinpoint the need for additional accuracy.