Section: New Results

Benchmarking

In modern High Performance Computing architectures, the memory subsystem is a common performance bottleneck. When optimizing an application, the developer has to study its memory access patterns and adapt accordingly the algorithms and data structures it uses. The objective is twofold: on one hand, it is necessary to avoid missuses of the memory hierarchy such as false sharing of cache lines or contention in a NUMA interconnect. On the other hand, it is essential to take advantage of the various cache levels and the memory hardware prefetcher. Still, most profiling tools focus on CPU metrics. The few of them able to provide an overview of the memory patterns involved by the execution rely on hardware instrumentation mechanisms and have two drawbacks. The first one is that they are based on sampling which precision is limited by hardware capabilities. The second one is that they trace a subset of all the memory accesses, usually the most frequent, without information ab out the other ones. In [30] we present Moca, an efficient tool for the collection of complete spatio-temporal memory traces. Moca is based on a Linux kernel module and provides a coarse grained trace of a superset of all the memory accesses performed by an application over its addressing space during the time of its execution. The overhead of Moca is reasonable when taking into account the fact that it is able to collect complete traces which are also more precise than the ones collected by comparable tools.

Benchmarking has proven to be crucial for the investigation of the behavior and performances of a system. However, the choice of relevant benchmarks still remains a challenge. To help the process of comparing and choosing among benchmarks, in [33] we propose a solution for automatic benchmark profiling. It computes unified benchmark profiles reflecting benchmarks' duration, function repartition, stability, CPU efficiency, parallelization and memory usage. It identifies the needed system information for profile computation, collects it from execution traces and produces profiles through efficient and reproducible trace analysis treatments. The paper presents the design, implementation and the evaluation of the approach.