Performance Analysis of Congestion in Commodity Networks for Exascale Computing Platforms

• Good understanding of the TCP protocols and ability to recompile the linux kernel • Basics of experiment analysis with R is a plus 1 Context There is a continued need for higher compute performance: scientific grand challenges, engineering, geo-physics, bioinformatics, etc. Such studies used to be carried out on large ad hoc supercomputers, which, for economical reasons, were replaced by commodity clusters, i.e., sets of off-the-shelf computers interconnected by fast switches. Indeed, the technological advances driven by the home PC market have contributed to achieving high performance in commodity components. For decades, computer performance had doubled every 18 months merely by increasing the clock frequency of the processors. This trend stopped last decade for reason of electricity consumption and heat. Indeed, the computational power of a computer increases nearly sub-linearly with clock frequency while the energy consumption increases more than quadratically. As an answer to the power and heat challenges, processor constructors have increased the amount of computing units (or cores) per processor. Modern High Performance Computing (HPC) systems comprise thousands of nodes, each of them holding several multi-core processors. For example, one of the world fastest computers, the IBM Sequoia system Laurence Livermoor National Laboratory (USA), contains 9