FPGA Prototyping of Emerging Manycore Architectures for Parallel Programming Research using Formic Boards

Performance evaluation of parallel software and architectural exploration of innovative hardware support face a common challenge with emerging manycore platforms: they are limited by the slow running time and the low accuracy of software simulators. Manycore FPGA prototypes are difficult to build, but they offer great rewards. Software running on such prototypes runs orders of magnitude faster than current simulators. Moreover, researchers gain significant architectural insight during the modeling process. We use the Formic FPGA prototyping board [1], which specifically targets scalable and cost-efficient multi-board prototyping, to build and test a 64-board model of a 512-core, MicroBlaze-based, non-coherent hardware prototype with a full network-on-chip in a 3D-mesh topology. We expand the hardware architecture to include the ARM Versatile Express platforms and build a 520-core heterogeneous prototype of 8 Cortex-A9 cores and 512 MicroBlaze cores. We then develop an MPI library for the prototype and evaluate it extensively using several bare-metal and MPI benchmarks. We find that our processor prototype is highly scalable, models faithfully single-chip multicore architectures, and is a very efficient platform for parallel programming research, being 50,000 times faster than software simulation.

GPRM: a high performance programming framework for manycore processors

Processors with large numbers of cores are becoming commonplace. In order to utilise the available resources in such systems, the programming paradigm has to move towards increased parallelism. However, increased parallelism does not necessarily lead to better performance. Parallel programming models have to provide not only flexible ways of defining parallel tasks, but also efficient methods to manage the created tasks. Moreover, in a general-purpose system, applications residing in the system compete for the shared resources. Thread and task scheduling in such a multiprogrammed multithreaded environment is a significant challenge. In this thesis, we introduce a new task-based parallel reduction model, called the Glasgow Parallel Reduction Machine (GPRM). Our main objective is to provide high performance while maintaining ease of programming. GPRM supports native parallelism; it provides a modular way of expressing parallel tasks and the communication patterns between them. Compiling a GPRM program results in an Intermediate Representation (IR) containing useful information about tasks, their dependencies, as well as the initial mapping information. This compile-time information helps reduce the overhead of runtime task scheduling and is key to high performance. Generally speaking, the granularity and the number of tasks are major factors in achieving high performance. These factors are even more important in the case of GPRM, as it is highly dependent on tasks, rather than threads. We use three basic benchmarks to provide a detailed comparison of GPRM with Intel OpenMP, Cilk Plus, and Threading Building Blocks (TBB) on the Intel Xeon Phi, and with GNU OpenMP on the Tilera TILEPro64. GPRM shows superior performance in almost all cases, only by controlling the number of tasks. GPRM also provides a low-overhead mechanism, called “Global Sharing”, which improves performance in multiprogramming situations. We use OpenMP, as the most popular model for shared-memory parallel programming as the main GPRM competitor for solving three well-known problems on both platforms: LU factorisation of Sparse Matrices, Image Convolution, and Linked List Processing. We focus on proposing solutions that best fit into the GPRM’s model of execution. GPRM outperforms OpenMP in all cases on the TILEPro64. On the Xeon Phi, our solution for the LU Factorisation results in notable performance improvement for sparse matrices with large numbers of small blocks. We investigate the overhead of GPRM’s task creation and distribution for very short computations using the Image Convolution benchmark. We show that this overhead can be mitigated by combining smaller tasks into larger ones. As a result, GPRM can outperform OpenMP for convolving large 2D matrices on the Xeon Phi. Finally, we demonstrate that our parallel worksharing construct provides an efficient solution for Linked List processing and performs better than OpenMP implementations on the Xeon Phi. The results are very promising, as they verify that our parallel programming framework for manycore processors is flexible and scalable, and can provide high performance without sacrificing productivity.

Techniques and tools for parallelizing software

With the emergence of multicore and manycore processors, engineers must design and develop software in drastically new ways to benefit from the computational power of all cores. However, developing parallel software is much harder than sequential software because parallelism can't be abstracted away easily. Authors Hans Vandierendonck and Tom Mens provide an overview of technologies and tools to support developers in this complex and error-prone task. © 2012 IEEE.

Prefetching and Cache Management using Task Lifetimes

Task-based dataflow programming models and runtimes emerge as promising candidates for programming multicore and manycore architectures. These programming models analyze dynamically task dependencies at runtime and schedule independent tasks concurrently to the processing elements. In such models, cache locality, which is critical for performance, becomes more challenging in the presence of fine-grain tasks, and in architectures with many simple cores.

This paper presents a combined hardware-software approach to improve cache locality and offer better performance is terms of execution time and energy in the memory system. We propose the explicit bulk prefetcher (EBP) and epoch-based cache management (ECM) to help runtimes prefetch task data and guide the replacement decisions in caches. The runtimem software can use this hardware support to expose its internal knowledge about the tasks to the architecture and achieve more efficient task-based execution. Our combined scheme outperforms HW-only prefetchers and state-of-the-art replacement policies, improves performance by an average of 17%, generates on average 26% fewer L2 misses, and consumes on average 28% less energy in the components of the memory system.

Paraphrasing:Generating parallel programs using refactoring

Refactoring is the process of changing the structure of a program without changing its behaviour. Refactoring has so far only really been deployed effectively for sequential programs. However, with the increased availability of multicore (and, soon, manycore) systems, refactoring can play an important role in helping both expert and non-expert parallel programmers structure and implement their parallel programs. This paper describes the design of a new refactoring tool that is aimed at increasing the programmability of parallel systems. To motivate our design, we refactor a number of examples in C, C++ and Erlang into good parallel implementations, using a set of formal pattern rewrite rules. © 2013 Springer-Verlag Berlin Heidelberg.

Hybrid address spaces: A methodology for implementing scalable high-level programming models on non-coherent many-core architectures

This paper introduces hybrid address spaces as a fundamental design methodology for implementing scalable runtime systems on many-core architectures without hardware support for cache coherence. We use hybrid address spaces for an implementation of MapReduce, a programming model for large-scale data processing, and the implementation of a remote memory access (RMA) model. Both implementations are available on the Intel SCC and are portable to similar architectures. We present the design and implementation of HyMR, a MapReduce runtime system whereby different stages and the synchronization operations between them alternate between a distributed memory address space and a shared memory address space, to improve performance and scalability. We compare HyMR to a reference implementation and we find that HyMR improves performance by a factor of 1.71× over a set of representative MapReduce benchmarks. We also compare HyMR with Phoenix++, a state-of-art implementation for systems with hardware-managed cache coherence in terms of scalability and sustained to peak data processing bandwidth, where HyMR demon- strates improvements of a factor of 3.1× and 3.2× respectively. We further evaluate our hybrid remote memory access (HyRMA) programming model and assess its performance to be superior of that of message passing.

Power and Energy Implications of the Number of Threads Used on the Intel Xeon Phi

Energy consumption has become an important area of research of late. With the advent of new manycore processors, situations have arisen where not all the processors need to be active to reach an optimal relation between performance and energy usage. In this paper a study of the power and energy usage of a series of benchmarks, the PARSEC and the SPLASH- 2X Benchmark Suites, on the Intel Xeon Phi for different threads configurations, is presented. To carry out this study, a tool was designed to monitor and record the power usage in real time during execution time and afterwards to compare the r

Computing probable maximum loss in catastrophe reinsurance portfolios on multi-core and many-core architectures

In the reinsurance market, the risks natural catastrophes pose to portfolios of properties must be quantified, so that they can be priced, and insurance offered. The analysis of such risks at a portfolio level requires a simulation of up to 800 000 trials with an average of 1000 catastrophic events per trial. This is sufficient to capture risk for a global multi-peril reinsurance portfolio covering a range of perils including earthquake, hurricane, tornado, hail, severe thunderstorm, wind storm, storm surge and riverine flooding, and wildfire. Such simulations are both computation and data intensive, making the application of high-performance computing techniques desirable.

In this paper, we explore the design and implementation of portfolio risk analysis on both multi-core and many-core computing platforms. Given a portfolio of property catastrophe insurance treaties, key risk measures, such as probable maximum loss, are computed by taking both primary and secondary uncertainties into account. Primary uncertainty is associated with whether or not an event occurs in a simulated year, while secondary uncertainty captures the uncertainty in the level of loss due to the use of simplified physical models and limitations in the available data. A combination of fast lookup structures, multi-threading and careful hand tuning of numerical operations is required to achieve good performance. Experimental results are reported for multi-core processors and systems using NVIDIA graphics processing unit and Intel Phi many-core accelerators.

General-Purpose Computing on Graphic Processing Unit (GPGPU): metodi e tecnologie per lo sviluppo di applicazioni

Microprocessori basati su singolo processore (CPU), hanno visto una rapida crescita di performances ed un abbattimento dei costi per circa venti anni. Questi microprocessori hanno portato una potenza di calcolo nell’ordine del GFLOPS (Giga Floating Point Operation per Second) sui PC Desktop e centinaia di GFLOPS su clusters di server. Questa ascesa ha portato nuove funzionalità nei programmi, migliori interfacce utente e tanti altri vantaggi. Tuttavia questa crescita ha subito un brusco rallentamento nel 2003 a causa di consumi energetici sempre più elevati e problemi di dissipazione termica, che hanno impedito incrementi di frequenza di clock. I limiti fisici del silicio erano sempre più vicini. Per ovviare al problema i produttori di CPU (Central Processing Unit) hanno iniziato a progettare microprocessori multicore, scelta che ha avuto un impatto notevole sulla comunità degli sviluppatori, abituati a considerare il software come una serie di comandi sequenziali. Quindi i programmi che avevano sempre giovato di miglioramenti di prestazioni ad ogni nuova generazione di CPU, non hanno avuto incrementi di performance, in quanto essendo eseguiti su un solo core, non beneficiavano dell’intera potenza della CPU. Per sfruttare appieno la potenza delle nuove CPU la programmazione concorrente, precedentemente utilizzata solo su sistemi costosi o supercomputers, è diventata una pratica sempre più utilizzata dagli sviluppatori. Allo stesso tempo, l’industria videoludica ha conquistato una fetta di mercato notevole: solo nel 2013 verranno spesi quasi 100 miliardi di dollari fra hardware e software dedicati al gaming. Le software houses impegnate nello sviluppo di videogames, per rendere i loro titoli più accattivanti, puntano su motori grafici sempre più potenti e spesso scarsamente ottimizzati, rendendoli estremamente esosi in termini di performance. Per questo motivo i produttori di GPU (Graphic Processing Unit), specialmente nell’ultimo decennio, hanno dato vita ad una vera e propria rincorsa alle performances che li ha portati ad ottenere dei prodotti con capacità di calcolo vertiginose. Ma al contrario delle CPU che agli inizi del 2000 intrapresero la strada del multicore per continuare a favorire programmi sequenziali, le GPU sono diventate manycore, ovvero con centinaia e centinaia di piccoli cores che eseguono calcoli in parallelo. Questa immensa capacità di calcolo può essere utilizzata in altri campi applicativi? La risposta è si e l’obiettivo di questa tesi è proprio quello di constatare allo stato attuale, in che modo e con quale efficienza pùo un software generico, avvalersi dell’utilizzo della GPU invece della CPU.

Optimization of Pattern Matching Algorithms for Multi- and Many-Core Platforms

Image and video compression play a major role in the world today, allowing the storage and transmission of large multimedia content volumes. However, the processing of this information requires high computational resources, hence the improvement of the computational performance of these compression algorithms is very important. The Multidimensional Multiscale Parser (MMP) is a pattern-matching-based compression algorithm for multimedia contents, namely images, achieving high compression ratios, maintaining good image quality, Rodrigues et al. [2008]. However, in comparison with other existing algorithms, this algorithm takes some time to execute. Therefore, two parallel implementations for GPUs were proposed by Ribeiro [2016] and Silva [2015] in CUDA and OpenCL-GPU, respectively. In this dissertation, to complement the referred work, we propose two parallel versions that run the MMP algorithm in CPU: one resorting to OpenMP and another that converts the existing OpenCL-GPU into OpenCL-CPU. The proposed solutions are able to improve the computational performance of MMP by 3 and 2:7 , respectively. The High Efficiency Video Coding (HEVC/H.265) is the most recent standard for compression of image and video. Its impressive compression performance, makes it a target for many adaptations, particularly for holoscopic image/video processing (or light field). Some of the proposed modifications to encode this new multimedia content are based on geometry-based disparity compensations (SS), developed by Conti et al. [2014], and a Geometric Transformations (GT) module, proposed by Monteiro et al. [2015]. These compression algorithms for holoscopic images based on HEVC present an implementation of specific search for similar micro-images that is more efficient than the one performed by HEVC, but its implementation is considerably slower than HEVC. In order to enable better execution times, we choose to use the OpenCL API as the GPU enabling language in order to increase the module performance. With its most costly setting, we are able to reduce the GT module execution time from 6.9 days to less then 4 hours, effectively attaining a speedup of 45 .