Fast and Faithful Performance Prediction of MPI Applications: the HPL Case Study

Cornebize, Tom; Legrand, Arnaud; Heinrich, Franz

doi:10.1109/cluster.2019.8891011

Cited by 5 publications

(8 citation statements)

References 21 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…As illustrated in Figure 3 and explained in our previous work [5], accurate predictions require careful modeling of both spatial and temporal variability, as they appear to have a very strong effect on HPL performance. Somehow, this is expected since HPL is an iterative program that synchronizes through the broadcast of factorization panels.…”

Section: A First Validationmentioning

confidence: 85%

“…In this section, we present an overview of Simgrid/SMPI and of the modifications of HPL required to obtain a scalable simulation and a first validation of the simulations. The results of Sections 3.1-3.4 previously appeared in a conference publication [5] and are included here for completeness.…”

Section: Emulating Hpl With Simgrid/smpimentioning

confidence: 99%

“…Last, the rationale for using a half-normal distribution rather than a normal distribution stems from the natural positive skewness of compute kernel duration. This model is much more complex than the simple deterministic one used in Figure 2 but, as we will explain, this complexity is key to obtain good performance prediction [5]. However, all other kernels' duration (which represent a negligible fraction of the overall execution time) have been modeled with a simple deterministic and homogeneous model such as daxpy(N ) = αN + β.…”

Section: Emulating Hplmentioning

confidence: 99%

“…In a previous publication [5], we presented how an application like HPL can be emulated at a reasonable cost on a single commodity server to study scenarios similar to qualification runs of supercomputers for the Top500 ranking [6].…”

Section: Introductionmentioning

confidence: 99%

“…This article is organized as follows: Section 2 presents the main characteristics of the HPL application and provides information on how the runs are conducted on modern supercomputers. In Section 3, we briefly present the simulator we used for this work, SimGrid/SMPI, and the modifications of HPL that were required to obtain a scalable simulation and some initial validation results presented in an earlier work [5]. These results highlight the importance of mod- eling both spatial and temporal variability.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

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

Cornebize¹,

Legrand²

2021

Preprint

Self Cite

View full text Add to dashboard Cite

Finely tuning MPI applications and understanding the influence of key parameters (number of processes, granularity, collective operation algorithms, virtual topology, and process placement) is critical to obtain good performance on supercomputers. With the high consumption of running applications at scale, doing so solely to optimize their performance is particularly costly. Having inexpensive but faithful predictions of expected performance could be a great help for researchers and system administrators. The methodology we propose decouples the complexity of the platform, which is captured through statistical models of the performance of its main components (MPI communications, BLAS operations), from the complexity of adaptive applications by emulating the application and skipping regular non-MPI parts of the code. We demonstrate the capability of our method with High-Performance Linpack (HPL), the benchmark used to rank supercomputers in the TOP500, which requires careful tuning. We briefly present ( 1) how the open-source version of HPL can be slightly modified to allow a fast emulation on a single commodity server at the scale of a supercomputer. Then we present (2) an extensive (in)validation study that compares simulation with real experiments and demonstrates our ability to predict the performance of HPL within a few percent consistently. This study allows us to identify the main modeling pitfalls (e.g., spatial and temporal node variability or network heterogeneity and irregular behavior) that need to be considered. Last, we show (3) how our "surrogate" allows studying several subtle HPL parameter optimization problems while accounting for uncertainty on the platform.

show abstract

Section: A First Validationmentioning

confidence: 85%

Section: Emulating Hpl With Simgrid/smpimentioning

confidence: 99%

Section: Emulating Hplmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations