Exa-SofT : HPC software and tools

Exa-SofT : Overview

Exa-SofT is one of the targeted projects (PC2) of the NumPEx PEPR, France’s exploratory research programme preparing the software ecosystem for exascale computing.

The end of Dennard scaling and the slowdown of Moore’s law have pushed supercomputers toward massive parallelism and hardware heterogeneity: exascale compute nodes combine ordinary CPUs with accelerators (mostly GPUs of varied architectures), alongside increasingly complex memory systems and interconnects. Developing numerical libraries and applications that are efficient, portable and maintainable on such machines is no longer a mere software-engineering problem — it calls for disruptive research across the whole software stack.

Rather than incrementally patching programming models designed for homogeneous machines, Exa-SofT takes a holistic approach spanning several layers of the stack — programming models, compilers, runtime systems, communication libraries, numerical libraries, and performance and energy analysis tools. The goal is a coherent, tightly integrated and exascale-ready software suite in which the expression of algorithms is decoupled from low-level architectural details.

Main scientific challenges

Productivity Performance portability Heterogeneity Scalability and resilience Energy efficiency

The project runs from 2023 to 2029 and is coordinated by Inria.

Consortium & Workpackages

Exa-SofT brings together five core partners plus one industrial partner that joined in 2025.

Consortium

Inria

Coordinator

with University of Bordeaux, Bordeaux INP, University of Rennes, University of Strasbourg and University of Grenoble-Alpes. Teams: Avalon, Camus, Cash, Concace, Diverse, Polaris, Roma, Storm, Tadaam, Topal.

CEA

Maison de la Simulation and DAM teams.

CNRS

with University of Toulouse and Sorbonne University. Teams: APO and SEPIA (IRIT, Toulouse), PEQUAN (LIP6, Paris).

Télécom SudParis

with ENSIIE. ACMES team.

Université Paris-Saclay

ParSys team (LISN).

IFPEN

Industrial partner, joined 2025

Workpackages

WP1

Efficient and composable programming models

High-level approaches to abstract applications from the machine, ease porting and debugging, and improve code composability and separation of concerns.

WP2

Just-in-Time code optimization

Making runtime systems reactive and proactive to the software and hardware context by bringing together JIT compilation and runtime task management.

WP3

Runtime systems at exascale

Porting task-based runtimes to exascale: communication and workload-distribution scalability, deep memory hierarchies, and fault tolerance.

WP4

Portable, scalable numerical building blocks and software

Task-based dense, sparse and multilinear (tensor) algebra libraries, including approximate-computing and mixed-precision algorithms.

WP5

Performance analysis and prediction

A scalable tracing tool, post-mortem and on-the-fly analysis, and energy modelling to feed back into the runtime.

WP6

Energy profiling and control

Fine-grain power/energy measurement and models, energy-aware scheduling at node level, and energy-aware job scheduling at cluster level.

A management work-package (WP0) handles coordination, dissemination and co-design with the other NumPEx projects. WP5 and WP6 act as cross-cutting activities across the whole stack.

Results & Impact

Selected scientific results

Recursive tasking

WP3

A task model that refines the task graph at the right granularity — typically 10–30% faster on dense linear algebra.

More details

Scalable, portable LU factorization with partial pivoting

WP4

Task-based HPL without pivoting synchronizations — about 97% of unpivoted performance at scale.

More details

Fine-grain energy measurement

WP5WP6

Per-task energy estimation below power-meter resolution, enabling energy-aware scheduling.

More details

Pallas trace format

WP5

A structured trace format analyzing large-scale traces in seconds instead of hours.

More details

Chameleon evolutions

WP4

A modern C++ interface aligned with std::linalg, extended to tensor operations, on StarPU.

More details

Approximate and mixed-precision computing

WP4

Randomized mixed-precision low-rank algorithms cutting cost and memory at fixed accuracy.

More details

Programming models

WP1WP2

PolyKokkos, STF++ and the COMET/PRISM composition and data-partitioning frameworks.

More details

Fault tolerance

WP3

Transparent low-overhead checkpoint/restart integrated into StarPU and NewMadeleine.

More details

Impact and outreach

Contribution to the launch of the Coordination Support Action “SPE” (Strategy for Post-Exascale) and to the ETP4HPC Strategic Research Agenda.
Prize of Innovation of the Académie des Sciences awarded to Samuel Thibault for StarPU (WP3).
New academic and industrial links: IFPEN, the ExaNLA initiative (Jülich), and collaborations with NVIDIA, Bull/Eviden and ATOS.
Involvement in defining the software environment of the coming French exascale machines (Jules Verne, Alice Recoque).
Sustained dissemination through journal and conference publications and community events (COMPAS, IPDPS, Cluster, InPEx, MUMPS Workshop), plus STEM outreach using EasyPAP.