Olivier Glück

Segment routing based traffic engineering for energy efficient backbone networks

Energy consumption has become a limiting factor for deploying large-scale distributed infrastructures. This work1 seeks to improve the energy efficiency of backbone networks by providing an intra-domain Software Defined Network (SDN) approach to selectively turn off a subset of links. We propose the STREETE framework (SegmenT Routing based Energy Efficient Traffic Engineering) that dynamically adapts the number of powered-on links to the traffic load. The core of the solution relies on SPRING, a novel protocol being standardized by IETF. It is also known under the name of Segment Routing. The algorithms have been implemented and evaluated using the OMNET++ simulator. Experimental results show that the consumption of 44% of links can be reduced while preserving good quality of service.

Comparison and tuning of MPI implementations in a grid context

Today, clusters are often interconnected by long distance networks to compose grids and to provide users with a huge number of available resources. To write parallel applications, developers are generally using the standard communication library MPI, which has been optimized for clusters. However, two main features of grids-long distance networks and technological heterogeneity-raise the question of MPI efficiency in grids. This paper presents an evaluation and tuning of four recent MPI implementations (MPICH2, MPICH-Madeleine, OpenMPI and YAMPII) in a research grid: Gridpsila5000. The comparison is based on the execution of pingpong and NAS parallel benchmarks. We show that these implementations present several performance differences. We show that YAMPII performs better results than the others. But we argue that executing MPI applications on a grid can be beneficial if some specific parameters are well tuned. The paper details, for each implementation, the tuning leading the best performances.

Optimizations of client's side communications in a distributed file system within a Myrinet cluster

High performance applications running on high-speed interconnects require both efficient communication between computing nodes and fast access to the storage system. Making the most of these networks to access remote files requires a good interaction between their highly specific software interface and the special requirements of distributed file systems. We show how non-buffered and buffered remote file access may be improved by modifying the network programming interface and firmware and by adding the required infrastructure in the operating system. Our modifications in Myrinet/GM show no performance penalty while the network usage in our ORFA (optimized remote file access) protocol is improved.

Energy efficiency in high-performance computing with and without knowledge of applications and services

The constant demand of raw performance in high-performance computing (HPC) often leads to over-provisioning in high-performance systems which in turn can result in a colossal energy waste due to workload/application variation over time. Proposing energy efficient solutions in the context of large-scale HPC is a real, unavoidable challenge. This article explores two alternative approaches (with or without knowledge of applications and services) dealing with the same goal: reducing the energy usage of large-scale infrastructures which support HPC applications. This article describes the first approach, with knowledge of applications and services, which enables users to choose the less consuming implementation of services. Based on the energy consumption estimation of the different implementations (protocols) for each service, this approach is validated on the case of fault tolerance service in HPC. The ‘without knowledge’ approach allows some intelligent framework to observe the life of HPC systems and proposes some energy reduction schemes. This framework automatically estimates the energy consumption of the HPC system in order to apply power saving schemes. Both approaches are experimentally evaluated and analysed in terms of energy efficiency.

An Efficient Network API for in-Kernel Applications in Clusters

Running parallel applications on clusters with highspeed local networks requires fast communication between computing nodes but also low latency and high bandwidth file access. However, the application programming interfaces of high-speed local networks were designed for MPI communication and do not always meet the requirements of other applications like distributed file systems. In this paper, we explore several solutions to improve the use of high-speed network for in-kernel applications. Distributed file systems implemented on top of the GM interface of MYRINET are first examined to demonstrate how hard it is to get an efficient interaction between such applications and the network. Then, we propose solutions to simplify and improve this interaction and integrate them into the kernel interface of the new MYRINET driver, MX. Performance comparisons between MX and GM, and their usage in both a distributed file system and a zero-copy protocol show nice improvements. Moreover, we are able to improve the performance of the flexible kernel API we designed in MX that allows to remove some intermediate copy

Energy-Aware Checkpointing Strategies

Future extreme-scale supercomputers will gather several millions of cores. The main problem that we address in this chapter is the energy consumption of these systems. Fault-tolerant methods must be deployed in such extreme-scale systems and these methods have a dramatic impact on total energy consumption. Fault-tolerant protocols have different energy consumption rates, depending on parameters such as platform characteristics, application features, and number of processes used in the execution. Currently, in order to evaluate the power consumption of fault-tolerant protocols in a given execution context, the only approach is to run the application with the different versions of fault-tolerant protocols and to monitor energy consumption. In order to avoid this time and energy consuming process, we describe in this chapter a methodology to estimate the energy consumption of the fault-tolerant protocols used for HPC applications. This methodology relies on an energy calibration of the supercomputer and a user description of the execution setting. We evaluate the accuracy of the estimations with applications and scenarios executed on a real platform with energy consumption monitoring. Results show that the energy estimations provided before the execution are highly accurate, and allow users to select the less energy consuming fault-tolerant protocol without pre-running their applications.

Providing Green Services in HPC Data Centers: A Methodology Based on Energy Estimation

A supercomputer is an infrastructure built from an interconnection of computers capable of performing tasks in parallel in order to achieve very high performance. They are used in order to run scientific applications in various fields like the prediction of severe weather phenomena and seismic waves. To meet new scientific challenges, the HPC community has set a new performance objective for the end of the decade: Exascale. To achieve such performance (1018 FLoat Operations Per Second), an exascale supercomputer will gather several millions of CPU cores running up to a billion trends and will consume several megawatts. The energy consumption issue at the exascale becomes even more worrying when we know that we already reach energy consumptions higher than 17 MW at the petascale while the DARPA set to 20 MW the threshold for exascale supercomputers. Hence, these systems that will be 30 times more performant than the current systems have to achieve an energy efficiency of 50 gigaFLOPS per watt while the current ones achieve between 2 and 3 gigaFLOPS per watt. As a consequence, reducing the energy consumption of high-performance computing infrastructures is a major challenge for the next years in order to be able to move to the exascale era.

Splitting TCP for MPI Applications Executed on Grids

In this paper, we first study the interaction between MPI applications and TCP on grids. Then, we propose MPI5000, a transparent applicative layer between MPI and TCP, using proxies to improve the execution of MPI applications on a grid. Proxies aim at splitting TCP connections in order to detect losses faster and avoid to return in a slow-start phase after an idle time. Finally, we evaluate our layer executing the NAS Parallel Benchmarks on Grid5000, the French research grid, using MPICH2. The results show that our architecture reduces the number of idle timeout and of long-distance retransmissions for BT, SP and LU benchmarks. Using MPI5000, these applications can decrease their execution time by 35%, 28%, and, 15% respectively. A comparison with MPICH-G2 performances shows that our layer can even outperform a grid enabled MPI implementation.