Search

This paper presents and analyzes a heterogeneous implementation of an industrial use case based on K-means that targets symmetric multiprocessing (SMP), GPUs and FPGAs. We present how the application can be optimized from an algorithmic point of view and how this optimization performs on two heterogeneous platforms. The presented implementation relies on the OmpSs programming model, which introduces a simplified pragma-based syntax for the communication between the main processor and the accelerators. Performance improvement can be achieved by the programmer explicitly specifying the data memory accesses or copies. As expected, the newer SMP+GPU system studied is more powerful than the older SMP+FPGA system. However the latter is enough to fulfill the requirements of our use case and we show that uses less energy when considering only the active power of the execution. This work is partially supported by the European Union H2020 project AXIOM (grant agreement n. 645496), HiPEAC (grant agreement n. 687698), and Mont-Blanc (grant agreements n. 288777, 610402 and 671697), the Spanish Government Programa Severo Ochoa (SEV-2015-0493), the Spanish Ministry of Science and Technology (TIN2015- 65316-P) and the Departament d’Innovació, Universitats i Empresa de la Generalitat de Catalunya, under project MPEXPAR: Models de Programaci´o i Entorns d’Execució Paral·lels (2014-SGR-1051).

In this thesis, I present a methodology to evaluate High-Performance Computing systems. The method relies on measurements at three levels: architectural features through micro-benchmarks; system software and tools through benchmarks and analysis of the code generated by the compiler; and sustained performance though scientific applications. I apply the method to three state-of-the-art High-Performance Computing clusters deployed at the Barcelona Supercomputing Center.

This project is about porting and tuning the Mont-Blanc benchmarks to the multicore ARM 64 bits architecture. The Mont-Blanc benchmarks are part of the Mont-Blanc European project and they have been developed internally in the BSC (Barcelona Supercomputing Center). The project will explore the possibilities that an ARM architecture can offer running in a HPC (High Performance Computing) setup, this includes to learn how to tune and adapt a parallelized computer program and analyze its execution behavior. As part of the project, we will analyze the performance of each benchmark using instrumentation tools such like Extrae and Paraver. Each benchmark will be adapted, tuned and executed mainly in the three new Mont-Blanc mini-clusters, Thunder (ARMv8 custom), Merlin (ARMv8 custom) and Jetson TX (ARMv8 cortex-a57) using the OmpSs programming model. The evolution of the performance obtained will be shown followed by a brief analysis of the results after each optimization. Aquest projecte es basa en adaptar i afinar els Mont-Blanc benchmarks a l’arquitectura multinucli ARM 64 bits. Els Mont-Blanc benchmarks formen part del projecte Europeu Mont-Blanc i han estat desenvolupats internament en el BSC (Barcelona Supercomputing Center). Aquest projecte explorarà el potencial d’usar l’arquitectura ARM en un entorn HPC (High Performance Computing), això inclou aprendre a adaptar i afinar un programa paral·lel, i analitzar el seu comportament durant l’execució. Com a part del projecte, s’analitzarà el rendiment de cada benchmark usant eines d’instrumentació com Extrae o Paraver. Cada benchmark serà adaptat, afinat i executat en els tres nous miniclústers de Mont-Blanc, Thunder (ARMv8 personalitzat), Merlin (ARMv8 personalitzat) i Jetson TX (ARMv8 cortex-a57) usant el model de programació OmpSs. Es mostrarà l’evolució del rendiment, seguit d’una breu explicació dels resultats després de cada optimització. Este proyecto se basa en adaptar y afinar los Mont-blanc benchmarks a la arquitectura multi-núcleo ARM 64 bits. Los Mont-Blanc benchmarks forman parte del proyecto Europeo Mont-Blanc y han sido desarrollados internamente en el BSC (Barcelona Supercomputing Center). Este proyecto explorará el potencial de usar la arquitectura ARM en un entorno HPC (High Performance Computing), esto incluye aprender a adaptar y afinar un programa paralelo, y analizar su comportamiento durante la ejecución. Como parte del proyecto, se analizará el rendimiento de cada benchmark usando herramientas de instrumentación como Extrae o Paraver. Cada benchmark será adaptado, afinado y ejecutado en los tres nuevos mini-clústeres de Mont-Blanc, Thunder (ARMv8 personalizado), Merlin (ARMv8 personalizado) y Jetson TX (ARMv8 cortex-a57) usando el modelo de programación OmpSs. Se mostrará la evolución del rendimiento obtenido, y una breve explicación de los resultados después de cada optimización.

High Performance Computing (HPC) is used for running advanced application programs efficiently, reliably, and quickly. In earlier decades, performance analysis of HPC applications was evaluated based on speed, scalability of threads, memory hierarchy. Now, it is essential to consider the energy or the power consumed by the system while executing an application. In fact, the High Power Consumption (HPC) is one of biggest problems for the High Performance Computing (HPC) community and one of the major obstacles for exascale systems design. The new generations of HPC systems intend to achieve exaflop performances and will demand even more energy to processing and cooling. Nowadays, the growth of HPC systems is limited by energy issues Recently, many research centers have focused the attention on doing an automatic tuning of HPC applications which require a wide study of HPC applications in terms of power efficiency. In this context, this paper aims to propose the study of an oceanographic application, named OceanVar, that implements Domain Decomposition based 4D Variational model (DD-4DVar), one of the most commonly used HPC applications, going to evaluate not only the classic aspects of performance but also aspects related to power efficiency in different case of studies. These work were realized at Bsc (Barcelona Supercomputing Center), Spain within the Mont-Blanc project, performing the test first on HCA server with Intel technology and then on a mini-cluster Thunder with ARM technology. In this work of thesis it was initially explained the concept of assimilation date, the context in which it is developed, and a brief description of the mathematical model 4DVAR. After this problem’s close examination, it was performed a porting from Matlab description of the problem of data-assimilation to its sequential version in C language. Secondly, after identifying the most onerous computational kernels in order of time, it has been developed a parallel version of the application with a parallel multiprocessor programming style, using the MPI (Message Passing Interface) protocol. The experiments results, in terms of performance, have shown that, in the case of running on HCA server, an Intel architecture, values of efficiency of the two most onerous functions obtained, growing the number of process, are approximately equal to 80%. In the case of running on ARM architecture, specifically on Thunder mini-cluster, instead, the trend obtained is labeled as "SuperLinear Speedup" and, in our case, it can be explained by a more efficient use of resources (cache memory access) compared with the sequential case. In the second part of this paper was presented an analysis of the some issues of this application that has impact in the energy efficiency. After a brief discussion about the energy consumption characteristics of the Thunder chip in technological landscape, through the use of a power consumption detector, the Yokogawa Power Meter, values of energy consumption of mini-cluster Thunder were evaluated in order to determine an overview on the power-to-solution of this application to use as the basic standard for successive analysis with other parallel styles. Finally, a comprehensive performance evaluation, targeted to estimate the goodness of MPI parallelization, is conducted using a suitable performance tool named Paraver, developed by BSC. Paraver is such a performance analysis and visualisation tool which can be used to analyse MPI, threaded or mixed mode programmes and represents the key to perform a parallel profiling and to optimise the code for High Performance Computing. A set of graphical representation of these statistics make it easy for a developer to identify performance problems. Some of the problems that can be easily identified are load imbalanced decompositions, excessive communication overheads and poor average floating operations per second achieved. Paraver can also report statistics based on hardware counters, which are provided by the underlying hardware. This project aimed to use Paraver configuration files to allow certain metrics to be analysed for this application. To explain in some way the performance trend obtained in the case of analysis on the mini-cluster Thunder, the tracks were extracted from various case of studies and the results achieved is what expected, that is a drastic drop of cache misses by the case ppn (process per node) = 1 to case ppn = 16. This in some way explains a more efficient use of cluster resources with an increase of the number of processes.

The thesis evaluates the current state-of-the-art of RISC architectures in HPC. Studying the performance, power, and energy to solution in heterogeneous SoCs. For the evaluation 2 arm platforms (CPU+GPU, CPU+FPGA), 1 RISC-V platform and 1 Open Source RISC-V core running in an FPGA have been tested.