ROOM 3. HPL for the global south

Tutorial. HPL for the global south

ROOM 3.

Moday September 18, 2023

08:00 to 12:00 and  14:00 to 17:00 hours

Instructors:

  • Lic. Marcos Mazzini, CCAD - UNC
  • Dr. Nicolás Wolovick, Argentina, CCAD - UNC

Program:

Part 1: HPL Benchmark in the Latin American Context

  • Introduction to HPL and its inconsistent status in HPC.
  • Overview of the HPC landscape in Latin America and the Global South.
  • Rmax and HPC hardware description in the region.
  • Discussion on HPL's relevance and challenges in the context of limited resources and stress testing.
  • Importance of HPL for scientific and computational research in the region.

Part 2: Hands-On with HPL on a Cluster

  • Calculating Rpeak: Understanding cluster max performance capabilities.
  • Initial compilation of HPL: Setting up a base configuration with Spack.
  • Description of HPL input parameters: Initial input file and tweak points.
  • Initial execution of HPL in the cluster's queue system: Observing system behavior of CPU, memory and networking.
  • Identifying optimization opportunities: iterative compilations and executions selecting:
    • The right compiler.
    • Compiler flags.
    • Libraries.
    • Target architecture.
    • Which combination to use in other cases.
    • Networking issues.

 

Chair(s):
Information

This tutorial will bring a situated view (south cone, global south, developing countries) of the HPL benchmark's peculiar position in the world of High-Performance Computing (HPC).

Despite its dominance as TOP500 list benchmark, HPL is criticized as a stunt benchmark that fails to represent real scientific workloads.

In the context of the Global South, where HPC technology is still emerging with occasional success and limited resources, this tutorial seeks to address two critical questions:

First, why is HPL important when the region often struggles to achieve even a fraction of the latest TOP500 entries.

Second, how can we optimize and obtain meaningful insights from this benchmark.

The tutorial covers topics from the HPC landscape in the Global South, to HPL's general description and details on how to run it effectively on different architectures.

Ultimately, the discussion revolves around understanding the significance of HPL for the HPC communities in the context of developing regions.

Participants will solve practical exercises and gain hands-on experience with HPL on a cluster setup. The tutorial will empower them to comprehend the benchmark's significance within the Latin American HPC context while mastering essential techniques for maximizing its potential on various architectures.

 

Student's prerequisites 

Linux user, basic console usage, basic math knowledge for matrix operations.

Special conditions for access the tutorial

Laptop with ssh-client and ssh key pair to access CCAD-UNC resources.


 

 

References:

https://spack.readthedocs.io/en/latest/

https://www.advancedclustering.com/act_kb/tune-hpl-dat-file/

https://netlib.org/benchmark/hpl/tuning.html

Instructor(s):
Nicolás Wolovick