sOMP: Simulating OpenMP Task-Based Applications with NUMA Effects

Abstract

Anticipating the behavior of applications, studying, and designing algorithms are some of the most important purposes for the performance and correction studies about simulations and applications relating to intensive computing. Often studies that evaluate performance on a single-node of a simulation don't consider Non-Uniform Memory Access (NUMA) as having a critical effect. This work focuses on accurately predicting the performance of task-based OpenMP applications from traces collected through the OMPT interface. We first introduce TiKKi, a tool that records a rich high-level representation of the execution trace of a real OpenMP application. With this trace, an accurate prediction of the execution time is modeled from the architecture of the machine and sOMP, a SimGrid-based simulator for task-based applications with data dependencies. These predictions are improved when the model takes into account memory transfers. We show that good precision (10% relative error on average) can be obtained for various grains and on different numbers of cores inside different shared-memory architectures.

Improving Simulations of Task-Based Applications on Complex NUMA Architectures

“…Hence, our previous Intel platform model [13] was insufficient, prompting extension for this study. Figure 1 shows our proposed SimGrid model for the AMD platform.…”

“…Our previous work [13] presented the sOMP simulator, which leverages the SimGrid framework to simulate task-based application execution with NUMA effects. Despite good predictions for Cholesky factorization, the project had several limitations:…”

“…sOMP The simulator, presented in our previous work [13] and tailored for taskbased applications with data dependencies, predicts their performance on architectures modeled in the SimGrid XML format using native execution-generated trace files. After parsing these files, sOMP submits tasks to a queue managed by a scheduler, akin to a standard OpenMP runtime.…”

“…An overview of NUMA architectures was presented in our previous work [13] for an Intel platform: specifically, an Intel Xeon Gold 6240 with 36 cores (Cascade-Lake microarchitecture) and two NUMA nodes, 18 cores each. The latter is represented with SimGrid components: the cores and the memory controllers are modeled as SimGrid hosts, the intrasocket interconnect is modeled as a SimGrid backbone, and the intersocket UPI link is modeled as a SimGrid link between routers.…”

“…Our earlier work [13] introduced a preliminary simulator, named sOMP, using SimGrid [9] to predict the performance of task-based applications on sharedmemory architectures by modeling a simple NUMA structure and data locality impacts. This tool allowed the non-cycle-accurate simulation of task-based applications using a trace of their sequential execution obtained with OMPT.…”