Simulation-based Optimization and Sensibility Analysis of MPI Applications: Variability Matters

TL;DR

This paper introduces a simulation-based methodology for predicting MPI application performance, enabling efficient tuning and sensitivity analysis by modeling key parameters and platform variability, demonstrated on the HPL benchmark.

Contribution

It presents a novel surrogate modeling approach that decouples platform complexity from application behavior, allowing accurate performance predictions and optimization insights for MPI applications.

Findings

Performance predictions within a few percent accuracy

Effective identification of modeling pitfalls like variability and heterogeneity

Enabling sensitivity analysis under platform uncertainty

Abstract

Finely tuning MPI applications and understanding the influence of keyparameters (number of processes, granularity, collective operationalgorithms, virtual topology, and process placement) is critical toobtain good performance on supercomputers. With the high consumptionof running applications at scale, doing so solely to optimize theirperformance is particularly costly. Havinginexpensive but faithful predictions of expected performance could bea great help for researchers and system administrators. Themethodology we propose decouples the complexity of the platform, whichis captured through statistical models of the performance of its maincomponents (MPI communications, BLAS operations), from the complexityof adaptive applications by emulating the application and skippingregular non-MPI parts of the code. We demonstrate the capability of our method with High-PerformanceLinpack (HPL), the benchmark used to rank supercomputers in theTOP500, which requires careful tuning. We briefly present (1) how theopen-source version of HPL can be slightly modified to allow a fastemulation on a single commodity server at the scale of asupercomputer. Then we present (2) an extensive (in)validation studythat compares simulation with real experiments and demonstrates our ability to predict theperformance of HPL within a few percent consistently. This study allows us toidentify the main modeling pitfalls (e.g., spatial and temporal nodevariability or network heterogeneity and irregular behavior) that needto be considered. Last, we show (3) how our "surrogate" allowsstudying several subtle HPL parameter optimization problems whileaccounting for uncertainty on the platform.