link to homepage

Institute for Advanced Simulation (IAS)

Navigation and service

HPC Optimisation and Scaling Workshop

To help make the most effective use of JSC supercomputing systems, the HPC Optimisation and Scaling Workshop from 18 to 22 February combined training from Intel and JSC in the use of their parallel application performance analysis tools with the opportunity to execute large-scale runs and measurements to validate correctness and investigate and improve scaling. Over twenty participants (often working in teams of two to four) were supported by a comparable number of JSC employees from the Simulation Laboratories and Cross-Sectional Teams, three instructors from Intel, and additional HPC systems support staff assisting in the background. JUWELS and JURECA (cluster and booster parts) had reservations each day to run executions with large numbers of compute nodes. These were progressively released when no longer required to minimize service disruption.

Participants were encouraged to provide their codes with test cases and/or execution measurements for auditing in advance of the workshop. Preliminary analyses of several codes were briefly presented at the start of the workshop and incorporated into plans developed during the workshop. Assessments are continuing in most cases.

Three codes which effectively employed multi-threading combined with MPI were readily scaled to run successfully with up to 2048 compute nodes of JUWELS, and two codes using only MPI (and no threading) were tuned to use 1024 compute nodes. Large-scale modular supercomputing combining JURECA cluster and booster nodes was also demonstrated. Other code teams focused on optimizing memory access and vectorization. The workshop participants were very grateful for the opportunity and assistance provided.

In the course of this thorough exercising of system hardware and software, a variety of issues were encountered, not all of which could be resolved during the workshop but are still actively being addressed. Usage of the supercomputing systems will thereby be improved, allowing the solution of larger computational problems and a more efficient use of resources.

Contact: Dr. Brian Wylie,

from JSC News No. 264, 27 March 2019