Brice Goglin and Samuel Thibault: HPC mapping experts

Imagine looking at a world map in which no city names or country borders appear and being asked to locate all of the natural resources available for a given activity. It would be a difficult task! But this is what happens with high-performance computing (HPC) infrastructure, as well as ordinary but increasingly complex PCs and servers, according to Brice Goglin and Samuel Thibault, creators of the hwloc (Portable Hardware Locality) software. 

“The origins of hwloc software date back to my thesis, and it was created to make it easier to identify and locate each of a supercomputer’s available resources (processor cores, cache, peripheral devices, etc.”, explains Samuel Thibault, a researcher working with Inria’s Storm project team (static optimization and execution methods). “This preliminary description of the hardware architecture allows better distribution of a program’s simultaneous computing tasks.” 

Open-source software that has become a standard

With this portable open-source software (which runs on any hardware) provided by the Bordeaux University Inria Centre since 2009, Samuel Thibault develops features that analyse the structure of what is to be computed and determine the best possible allocation of tasks. His current research concerns mainly ‘run-time support’, i.e., optimization software intended for concurrent applications (those that run different tasks simultaneously with the help of multiple processors). 

This work is completed by that of Brice Goglin, the TADaaM  team Director of Research who is a network, memory and parallel architecture specialist. “Machines can be complex, often with several processors and multiple cores”, the researcher emphasises. “hwloc software provides an abstraction, hence ensuring that the solution is running optimally, without the need to analyse or understand the hardware characteristics. The software is independent of the hardware, which facilitates its use with a large number of solutions.”  And the fact that it is an open-source software is vital from a research viewpoint. “As a common asset, it can be downloaded and enhanced by both individuals and industrial players like Intel, AMD and Dell, who can ensure that their hardware is compatible with hwloc and any software that uses it”, sums up Samuel Thibault. 

Embedded in Frontier, the world’s most powerful supercomputer

The software was adopted very quickly by users across the planet. “Right from the start, hwloc was praised by Jeff Squyres, who works on high-performance communications at Cisco”, says Brice Goglin. “He added it to the Open MPI library, which was under his responsibility, and which comprises tools intended for speeding up the exchange of data relating to supercomputer processes. It then took off with a snowball effect. Nowadays, all MPI distributions use hwloc, as do all HPC software stacks in academia and in industry, as well as several Unix-derived solutions, such as Apache Traffic Server” (software that aims to enhance the performance of web servers).

Another important measure of its recognition is that since 2022, the software has been embedded in the American supercomputer, Frontier, thought to be the most powerful in the world. This gigantic concurrent computing machine has a capacity of 1 billion, billion operations per second, more than an exaFLOP. “To obtain results like that, every microsecond is crucial”, Brice Goglin points out. “The hwloc software, which we’ve adapted here to a processor tailor-made by AMD, has helped industrial users to find optimization solutions.” And this success is just waiting to be repeated... Currently, ongoing developments to improve its handling of hybrid processors, combining classic cores with more energy-efficient ones, as well as the handling of heterogeneous memories, combining classic memory with that generally associated with graphics processors, for example. The hwloc scientific venture, then, is far from over.