Section: Partnerships and Cooperations

International Initiatives

Inria International Labs

• JLESC (Joint Laboratory on Exascale Computing)

  The Corse team is involved in the JLESC with collaborations with UIUC (Sanjay Kalé) and BSC (Mont-Blanc projects). Kevin Pouget, Brice Videau and Jean-François Méhaut attended to the two JLESC workshops (Barcelona and Bonn) in 2015.

  • Energy Efficiency and Load Balancing

  • The power consumption of High Performance Computing (HPC) systems is an increasing concern as large-scale systems grow in size and, consequently, consume more energy. In response to this challenge, we propose new energy-aware load balancers that aim at reducing the energy consumption of parallel platforms running imbalanced scientific applications without degrading their performance. Our research explores dynamic load balancing, low power manycore platforms and DVFS techniques in order to reduce power consumption.

  • We propose the improvement of the performance and scalability of parallel seismic wave models through dynamic load balancing. These models suffer from load imbalance for two reasons. First, they add a specific numerical condition at the borders of the domain, in order to absorb the outgoing energy. The decomposition of the domain into a grid of subdomains, which are distributed among tasks, creates load differences between the tasks that simulate the borders and those responsible for the central subdomains. Second, the propagation of waves in the simulated area changes the workload on the subdomains on different time-steps. Therefore causing dynamic load imbalance. In order to evaluate the use of dynamic load balancing, we ported a seismic wave simulator to Adaptive MPI, to benefit from its load balancing framework. Our experimental results show that dynamic load balancers can adapt to load variations during the application’s execution and improve performance by 36%.

  • we also focus on reducing the energy consumption of imbalanced applications through a combination of load balancing and Dynamic Voltage and Frequency Scaling (DVFS). Our strategy employs an Energy Daemon Tool to gather power information and a load balancing module that benefits from the load balancing framework available in the CHARM++ runtime system. We propose two variants of our energy-aware load balancer (ENERGYLB) to save energy on imbalanced workloads without considerably impacting the overall system performance. The first one, called Fine- Grained EnergyLB (FG-ENERGYLB), is suitable for plat- forms composed of few tens of cores that allow per-core DVFS. The second one, called Coarse-Grained EnergyLB (CG-ENERGLB) is suitable for current HPC platforms composed of several multi-core processors that feature per-chip DVFS.

Inria Associate Teams Not Involved in an Inria International Labs

IOComplexity

• Title: Automatic characterization of data movement complexity

• International Partner (Institution - Laboratory - Researcher):

  • Ohio State University (United States) - P. Sadayappan

• Start year: 2015

• See also: https://team.inria.fr/corse/iocomplexity/

• The goal of this project is to develop new techniques and tools for the automatic characterization of the data movement complexity of an application. The expected contributions are both theoretical and practical, with the ambition of providing a fully automated approach to I/O complexity characterization, in starking contrast with all known previous work that are stricly limited to pen-and-paper analysis.

  I/O complexity becomes a critical factor due in large part to the increasing dominance of data movement over computation in energy consumption for current and emerging architectures. This project aims at enabling: 1. the selection of algorithms according to this new criteria (as opposed to the criteria on arithmetic complexity that has been used up to now); 2. the design of specific architectures in terms of cache size, memory bandwidth, GFlops etc. based on application-specific bounds on memory traffic; 3. higher quality feedback to the user, the compiler, or the run-time system about data traffic, a major performance and energy factor.

PROSPIEL

• Title: Profiling and specialization for locality

• International Partner (Institution - Laboratory - Researcher):

  • Universidade Federal de Minas Gerais (Brazil) - Computer Science Department - Fernando Magno Quintão Pereira

• Start year: 2015

• See also: https://team.inria.fr/alf/prospiel/

• The PROSPIEL project aims at optimizing parallel applications for high performance on new throughput-oriented architectures: GPUs and many-core processors. Traditionally, code optimization is driven by a program analysis performed either statically at compile-time, or dynamically at run-time. Static program analysis is fully reliable but often over-conservative. Dynamic analysis provides more accurate data, but faces strong execution time constraints and does not provide any guarantee. By combining profiling-guided specialization of parallel programs with runtime checks for correctness, PROSPIEL seeks to capture the advantages of both static analysis and dynamic analysis. The project relies on the polytope model, a mathematical representation for parallel loops, as a theoretical foundation. It focuses on analyzing and optimizing performance aspects that become increasingly critical on modern parallel computer architectures: locality and regularity.

Exase

• Title: Exascale Computing Scheduling Energy

• See also: https://team.inria.fr/exase/

• Inria leader: Jean-Marc Vincent (Mescal)

• Inria teams: Mescal, Moais, Corse

• Corse participants: Jean-François Méhaut, François Broquedis, Frédéric Desprez

• International Partner (Institution - Laboratory - Researcher):

  • Federal University of Rio Grande do Soul (UFRGS, Porto Alegre, Brazil) - Informatics Faculty - L. Schnoor, N. Maillard, P. Navaux

  • Pontifical University Minas (PUC Minas, Belo Horizonte, Brazil) - Computer Science faculty, Henrique Freitas

  • University of Sao Paulo (USP, Sao Paulo, Brazil), IME faculty, Alfredo Goldman

• Start year: 2014

• The main scientific goal of Exase for the three years is the development of state-of- the-art energy-aware scheduling algorithms for exascale systems. As previously stated, issues on energy are fundamental for next generation parallel platforms and all scheduling decisions must be aware of that. Another goal is the development of trace analysis techniques for the behavior analysis of schedulers and the applications running on exascale machines. We list below specific objectives for each development axis presented in the previous section. analysis.

  • Fundamentals for the scaling of schedulers

  • Design of schedulers for large-scale infrastructures

  • Tools for the analysys of large scale schedulers

Participation in Other International Programs

• LICIA (LIG, UFRGS Brazil)

• EnergySFE (STIC Amsud)

  • Leader: University Federal of Santa Catarina (UFSC): Màrcio Castro

  • Partners: UFSC (Florianapolis, Brazil), UFRGS (Porto Alegre, Brazil), ESPE (Ecuador), CNRS (LIG/Corse , TIMA, LSPSC)

  • Duration: January 2016 - December 2017

  • Corse participants: Jean-François Méhaut, François Broquedis, Frédéric Desprez

  • The main goal of the EnergySFE research project is to propose fast and scalable energy-aware scheduling and fault tolerance techniques and algorithms for large-scale highly parallel architectures. To achieve this goal, it will be crucial to answer the following research questions:

    • How to schedule tasks and threads that compete for resources with different constraints while considering the complex hierarchical organization of future Exascale supercomputers?

    • How to tolerate faults without incurring in too much overhead in future Exascale supercomputers?

    • How scheduling and fault tolerance approaches can be adapted to be energy-aware?

    The first EnergySFE workshop was organized by the Corse team a the Inria Minatec building in September 2016.