High Performance Computing Challenges and Trends Claude TADONKI Mines ParisTech – LAL / CNRS / INP2P3 University of Oujda (Morocco) – October 7, 2011 High Performance Computing: Challenges and Trends Claude TADONKI The need of competitive HPC systems Increasing Need of HPC (range of applications and computing power) Large-scale scientific & technical computing (numerical and non numerical) Large-scale data mining and statistics in experimental physics Image and signal processing Video and 3D animations Gaming Molecular biology and structural genomic Sorting and pattern matching Meteorology, Atmospheric studies, Medical research Scientific and technical simulation High-standard industrial activities, services, and research investigations And more … Key Factors Massively parrallel computers & systems Dedicated architectures Specialized processors Processor frequency High level integration Memory space, latency, and bandwidth High speed interconnection Advances in parallel algorithm synthesis and programming language features Powerful compilers University of Oujda (Morocco) – October 7, 2011 High Performance Computing: Challenges and Trends Claude TADONKI Observations Parallel computing easily justifies nowadays (processor frequency evolution) • Frequency has been multiplied by 10 since 1993 • The number of transistors in Intel proc has been multiplied by … 100 000 since 1971 • Core voltage has been reduced by 10 (1,2V , the min is 0,7V) However, this evolution is closed to its asymptotic threshold !!! HPC interest covers a larger spectra of applications, hence a wider audience Performance expectations in some specific areas are beyond the capacity of standard computers Research Grid (Oxford University/IBM/United Devices) Oustanding achievment in large-scale molecular Alternatives and trends Smallpox analysis. Multi-core processors TLP-based parallel machines GPU (assume a skillful use!) Hardcoded (embedded) solutions Reconfigurable HPC Grid/Meta Computing University of Oujda (Morocco) – October 7, 2011 High Performance Computing: Challenges and Trends Claude TADONKI The K Computer RIKEN / Fujitsu ( JAPAN ) Number one the 37th TOP500 8.162 petaflops (Linpack) 93% 68,544 energy-efficient CPUs 672 computer racks Complete deployment in 2012 10 petaflops expected (2012) 800 computer racks (2012) Worldwide shared use http://www.fujitsu.com/global/news/pr/archives/month/2011/20110620-02.html University of Oujda (Morocco) – October 7, 2011 High Performance Computing: Challenges and Trends Claude TADONKI The Sequoia System LLNL / IBM ( USA ) BlueGene technology 500 teraFLOPS 1.6 million of cores 96 computer racks https://asc.llnl.gov/computing_resources/sequoia/ 98,304 compute nodes 1.6 petabytes of memory 10 times faster than today’s Most powerful system Sequoia in 1 hour = 6.7 billion people calculating 24h/24h during 320 years University of Oujda (Morocco) – October 7, 2011 High Performance Computing: Challenges and Trends About sustained performance Claude TADONKI Even an optimal algorithm will run at a fraction of the peak performance !!! Before we calculate, we need data Time-to-CPU could be long % flops Memory space decreases with level Shared use of memory slowdown Synchronization and data exchange Control flow (if, while, for, case, …) University of Oujda (Morocco) – October 7, 2011 High Performance Computing: Challenges and Trends Important considerations Claude TADONKI vital on embedded systems Energy consumption and dissipation significant heat and cause of failure Integration ( more powerful unit/system in a smaller chip/surface ) Total cost of the system and maintenance issues Programmability (consider the case of the IBM CELL) Computing nodes interconnection (topology and speed) Accessibility (remote and shared use among several entities) Software and tools (system, programming, monitoring, profiling, …) Lifetime and evolution of the system (extensibility, devices change/upgrade, …) University of Oujda (Morocco) – October 7, 2011 High Performance Computing: Challenges and Trends Claude TADONKI Fundamental aspects Design of efficient parallel algorithms (modelling & scheduling) Complexity & Performance analysis Algorithmic and programming paradigms Code generation and transformations (automatic parallelization) Compilation techniques University of Oujda (Morocco) – October 7, 2011 High Performance Computing: Challenges and Trends Claude TADONKI Programming models Distributed memory parallel programming (MPI) Shared memory parallel programming (OpenMP, Pthreads) Accelerator-based programming (GPU, FPGA, CELL, …) Vector programming (SSE, VMX, SPU intrinsics, …) Hybrid programming (MPI+OpenMP/Pthread, CPU+GPU, PPU+SPU, …) University of Oujda (Morocco) – October 7, 2011 High Performance Computing: Challenges and Trends Claude TADONKI University of Oujda (Morocco) – October 7, 2011