Loop Block Profiling with Performance Prediction

With increase in the complexity of High Performance Computing systems, the complexity of applications has increased as well. To achieve better performance by effectively exploiting parallelism from High Performance Computing architectures, we need to analyze/identify various parameters such as, the code hotspot (kernel), execution time, etc of the program. Statistics say that a program usually spends 90% of the time in executing less than 10% of the code. If we could optimize even some small portion of the 10% of the code that takes 90% of the execution time we have a high probability of getting better performance. So we must find the bottleneck, that is the part of the code which takes a long time to run which is usually called the hotspot. Profiling provides a solution to the question: which portions of the code should be optimized/parallelized, for achieving better performance. In this research work we develop a light-weight profiler that gives information about which portions of the code is the hotspot and estimates the maximum speedup that could be achieved, if the hotspot is parallelized. Keywords—Profiling, Loop Block Profile, Code Analysis, Performance Prediction, Speedup Estimation