Section: New Results

HPC support for Aevol

participants: Jonathan Rouzaud-Cornabas, David Parsons, Guillaume Beslon

During the year, we had three internships that focus around HPC. The three of them were founded through the Federation Informatique de Lyon (FIL FR2000) and were common between the Inria Beagle team (LIRIS) and the Inria Avalon team (LIP).

The first one (Lukas Schmidt - M2) was working on component-based software engineering and HPC with Aevol as use-case. The goal was to see if and how the COMET [1] task-based parallel component model (and its implementation Halley) can fit the parallelization requirement of Aevol. An extension of the model was proposed to support hierarchical data structure and a prototype implementation has been done. In the future, we will work on the formalization of the extension and an efficient implementation on it. The goal is to ease the development and replacement of core components of the Aevol software (e.g. be able to easily replace the 2-base DNA code by a 4-base one).

The second internship (Valentin Huguet - M2) was evolving around Aevol and how to ease the distribution of the computation. To do so, an extension of the DIET software [2] was proposed and a fully functional webboard was implemented. We have a first prototype that support the execution of a large set of distributed computing resources and the control of its execution through a webboard. Moreover, basic visualization of the simulation results can be done through the same webboard. A following internship (starting Feb. 2019) will continue the work. The goal is to support workflow composed of multiple execution of Aevol and its pre/post treatments to automate the execution of large campaign that are done manually at the moment.

The goal of the third internship (Nathan Payre - L3) was to propose a prototype of a bitset for Aevol and its efficient implementation on modern hardware (Intel Skylake and Intel Xeon Phi). Indeed, the current implementation of Aevol DNA (2 base) uses a char type (8bit) to store a bit value (0 or 1). Accordingly, working at the bitset level could save up to 8 time memory space and speed up the computation (as Aevol is memory bound, reducing the memory transfer by 8 could dramatically speed up the global execution). Moreover, modern processors have vectorization extension that are perfectly fitting our requirements (we could process 512bit per cycle with AVX512 extension). During the internship, the bitset and the different operation we use in Aevol model (e.g. Hamming distance) were formalized and implemented. The preliminary results show a speed up of 140x on these operations. A full evaluation on the impact of the performance of Aevol and how different modern processor react to such implementation will be done in the future.

Last, a part of the Beagle team (Guillaume Beslon, David Parsons, Jonathan Rouzaud-Cornabas) were selected and participate to the EuroHack 2018 GPU Programming Hackathon in Lugano (Switzerland) organized by CSCS (Swiss National Supercomputing Centre) and NVidia. The goal was to port Aevol to modern GPU and thus to the CUDA programming language. In order to be able to do so in a week, we propose a mini-application (mini-Aevol) of Aevol [3] that is representative of the computation and memory pattern of the full Aevol. This prototype will be reuse in our collaboration with team focusing on HPC research. At the end of the week, we had a full implementation of mini-Aevol on GPU. New core algorithms of Aevol have been proposed to support massively parallel processors such as GPU. The prototype will be transfer to the full Aevol code in the future to be able to support GPU. It is worth noting that this mini-apps is also used in teaching context (INSA Lyon - Computer Science M2) to learn how to parallelize and optimize code with OpenMP and CUDA.

[1] Olivier Aumage, Julien Bigot, Hélène Coullon, Christian Pérez, Jérôme Richard. Combining Both a Component Model and a Task-based Model for HPC Applications: a Feasibility Study on GYSELA. 17th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGrid)., May 2017, Madrid, Spain.

[2] https://graal.ens-lyon.fr/diet/

[3] https://gitlab.inria.fr/rouzaudc/mini-aevol

[4] https://github.com/fomics/EuroHack18/blob/master/final/beagle_aevol.pdf