Using SimPoint for accurate and efficient simulation

Abstract

Modern architecture research relies heavily on detailed pipeline simulation. Simulating KeywordsSimPoint, Clustering, Simulation, Fast-forwarding, Sampling SIMPOINTUnderstanding the cycle level behavior of a processor running an application is crucial to modern computer architecture research. To gain this understanding, detailed cycle level simulators are typically employed. Unfortunately, this level of detail comes at the cost of speed, and simulating the full execution of an industry standard benchmark on even the fastest simulator can take weeks to months to complete. This fact has not gone unnoticed in the academic community, and several researchers have started to develop techniques aimed at reducing simulation time.For architecture research it is often necessary to take one instance of a program with a given input, and simulate its performance over many different architecture configurations. The same program binary with the input may be run hundreds or thousands of times to examine how, for example, the effectiveness of a given architecture changes with its size. Our goal in creating SimPoint [1, 2] is to (1) significantly reduce simulation time, (2) provide an accurate characterization of the full program, and (3) to perform the analysis to accomplish the first two goals in a matter of minutes. These goals are met by simulating only a handful of intelligently chosen sections of the full program. When these sections (simulation points) are carefully chosen, it provides an accurate picture of the complete execution of the program and results in highly accurate estimations of performanceThe key to our approach is that for a given program and input, the simulation points only need to be chosen once. This is because we select them using a method that is independent of any particular architecture configuration. The simulation points are selected using a metric that is only based on the code that is executed over time for a program/input pair. Once the sim- ulation points are chosen they can be used for the hundreds or thousands of independent simulations that may be needed, significantly reducing simulation time.To pick the simulation points in [1, 2], we introduce the concept of profiling Basic Block Vectors (BBV) as a way of capturing the important behaviors of the program over time [1]. A Basic Block Vector captures the relative frequency of the code blocks executed during a given portion of execution. After profiling a program with a particular input, we compare the basic block vectors to see how similar they are to one another. Intervals of execution that execute the same code blocks with the same frequency are grouped together into clusters using clustering algorithms from machine learning. We found that sections of execution (represented by basic block vectors) that are grouped into the same cluster have very similar behavior across all the architecture metrics we have examined. Once we break the program into clusters, we pick a single point from each cluster (appropriately weighted) to serve ...

ParaShares: Finding the Important Basic Blocks in Multithreaded Programs

“…The frequently executed blocks are called "hot" and are important optimization targets as they constitute a large share of an application's dynamic work. Hot block analysis has traditionally been used for a variety of purposes, including JIT translation [24], garbage collection optimizations [13], simulation points analysis [19], code cache management [20], and parallel performance debugging, for example, in Intel's VTune Amplifier [15].…”

The Smart Cache: An Energy-Efficient Cache Architecture Through Dynamic Adaptation

“…Since we start our experiments from the beginning of each application and execute 60 billion instructions without using any profiled phase information, many small transition phases are encountered in our experiments. These transition phase boundaries could be easily identified by a phase detector [26,23], which would allow us to vary the interval length and reconfigure more accurately. However, on average, our Smart cache approach achieves an energy-delay of 0.50, almost half that of the best static scheme.…”

Understanding transactional memory performance

“…While benchmark size and startup effects are a known problem for simulation [25], they continue to impede accurate measurement. We note that researchers can compensate for the brevity of the simulator inputs and still accurately calculate the gains in performance.…”