Quick facts
-
- Talence
Categories:
- Technology
Subjects:
- Phd
Apply by: 2023-07-25
PhD Position F/ M Topology-aware load balancing for ocean simulation on heterogeneous platforms.
Contexte et atouts du poste
Various HPC improvements of the CROCO model itself are currently carried out with respect to a sustainable support of GPUs and different parallel programming models. Indeed, the current trend in high-performance computing architectures is going even more towards increasing heterogeneity. This is omnipresent on the intra-node computation with accelerator cards as well as on the inter-node level with different hardware and communication behaviors.
However, on the application and scheduling side, this trend is often ignored: scheduling of applications, in particular CROCO, still assumes homogeneity across the hardware stack. This leads to a mismatch between applications and the underlying HPC system, resulting in a poor performance in particular in the strong scaling case.
The AIRSEA team in Grenoble is one of the main developers of the CROCO model and the Tadaam team in Bordeaux has the expertise in load-balancing and topology-aware algorithms. Therefore, this PhD will be carried out mainly in Bordeaux but with strong collaboration with Grenoble : visits and exchanges will be organized regularly between the two locations.
Mission confiée
The CROCO ocean model has a very complex workload model including non-homogeneous workload, adaptive mesh refinement with nested grids as well as existing support for hybrid CPUs and GPUs. Optimization attempts without application-driven information are therefore prone to fail. The goal of this PhD is to work on optimizing the execution of the CROCO model on supercomputers by developing and investigating new load-balancing algorithms.
Even if CROCO relies on structured meshes, load imbalance appears between the different computing units due to varying runtime of solvers. Moreover, as the topology of a heterogeneous machine can be extremely complex, the cost of communication can be very high depending on the location of the sender and the receiver. Hence, it is necessary to carefully optimize the mapping of the compute process and the load balance between them to optimize the computation and communication costs of the CRCOCO model.
Principales activités
The Phd Candidtae will work on the following workplan:
Compétences
Avantages
&v=024){kind=logo}
