Characterization of Multithreaded Scientific Workloads on Simultaneous Multithreading Intel Processors

Simultaneous Multithreading (SMT) is a technique that allows multiple independent threads to execute different instructions each cycle. Hyper-Threading (HT) is an implementation of SMT available on recent processors from Intel. Naturally, Multi-threaded applications are very suitable for SMT systems. However, HT due to extensive resource sharing may not suitably benefit OpenMP high performance computing applications. In this paper, we first present performance of different OpenMP constructs on dual and quad HT-based Intel Xeon servers under 2.4.22 and 2.6.9 kernels. We find that the overhead of OpenMP constructs with HT is an order of magnitude larger than when HT is off. We then use a range of OpenMP applications from the NAS and SPEC OMPM2001 suites to measure performance on a dual Hyper-Threaded SMP. Our performance results indicate majority of applications benefit from having a second thread in one-processor situations. However, only a few applications enjoy performance gain when HT is enabled on both processors. Data from hardware performance counters verifies trace cache misses and its delivery rate are sources of performance bottleneck.