Section: New Results

Energy Efficiency in Large Scale Systems

Participants : Ghislain Landry Tsafack, Mohammed El Mehdi Diouri, Olivier Glück, Laurent Lefevre.

Energy Efficiency in HPC Systems

Modern high performance computing subsystems (HPC) – including processor, network, memory, and I/O — are provided with power management mechanisms. These include dynamic speed scaling and dynamic resource sleeping. Understanding the behavioral patterns of high performance computing systems at runtime can lead to a multitude of optimization opportunities including controlling and limiting their energy usage. We have proposed a general purpose methodology for optimizing energy performance of HPC systems considering processor, disk and network. We have relied on the concept of execution vector along with a partial phase recognition technique for on-the-fly dynamic management without any a priori knowledge of the workload. We have demonstrated the effectiveness of our management policy under two real-life workloads. Experimental results have shown that our management policy in comparison with baseline unmanaged execution saves up to 24% of energy with less than 4% performance overhead for our real-life workloads [28] , [27] , [26] . This work is done under the Large Scale Initiative Hemera project (Joint PhD between Avalon and IRIT (Toulouse) with J.-M. Pierson, P. Stolf and G. Da Costa).

Energy Considerations in Checkpointing and Fault Tolerance Protocols

Two key points should be taken into account in future exascale systems: fault tolerance and energy consumption. To address these challenges, we evaluated checkpointing and existing fault tolerance protocols from an energy point of view. We measured on a real testbed the power consumption of the main atomic operations found in these protocols: checkpointing, message logging and coordination. The results [16] , [51] show that process coordination and RAM logging consume more power than checkpointing and HDD logging. However, the results we presented in Joules per Bytes for I/O operations, emphasize that checkpointing and HDD logging consume more energy than RAM logging because of the logging duration which is much more higher on HDD than on RAM . We have also shown that for identical nodes performing the same operation, the extra power cost due to this operation is the same. In general, we have learned that the power consumption of a node during a given operation remains constant during this operation. The power consumption of such a node is equal to its idle power consumption to which we add the extra power consumption due to the operation it is performing. Finally, we proposed to consider energy consumption as a criterion for the choice of fault tolerance protocols. In terms of energy consumption, we should promote message logging for applications exchanging small volumes of data and coordination for applications involving few processes. This work is a joint work with F. Cappello (Inria-UIUC-NCSA Joint Laboratory for Petascale Computing).

Towards a Smart and Energy-Aware Service-Oriented Manager for Extreme-Scale Applications

To address the issue of energy efficiency for exascale supercomputers, we proposed a smart and energy-aware service-oriented manager for exascale applications: SEASOMES [17] . This framework aggregates the various energy-efficient solutions to "consume less" energy and to "consume better". It involves both internal and external interactions with the various actors interfering directly or indirectly with the supercomputer. On the one hand, we recommended a more fine-grained collaboration between application and hardware resources in order to reduce energy consumption and provide sustainable exascale services. On the other hand, we suggested a cooperation between the user, the administrator, the resource manager and the energy supplier for the purpose of "consuming better".