作者: Weifeng Zhang , B. Calder , D.M. Tullsen
DOI: 10.1109/PACT.2005.7
关键词: Overhead (computing) 、 Program optimization 、 Profile-guided optimization 、 CAS latency 、 Parallel computing 、 Speedup 、 Computer science 、 Optimizing compiler 、 Dynamic programming 、 Multithreading
摘要: Dynamic optimization has the potential to adapt program's behavior at run-time deliver performance improvements over static optimization. systems usually perform their in series with application's execution. This incurs overhead which reduces benefit of dynamic optimization, and prevents some aggressive optimizations from being performed. In this paper we propose a new framework called Trident. Concurrent execution, uses hardware support identify opportunities, spare threads on multithreaded processor for these events. We evaluate using Trident guide code layout, basic compiler optimizations, value specialization. Our results show that achieves an average 20% speedup, is complementary other memory latency tolerant techniques, such as prefetching.