AMD Magny-Cours Research Articles

Parallel Discrete Event Simulation for Multi-Core Systems: Analysis and Optimization

Parallel Discrete Event Simulation (PDES) can substantially improve the performance and capacity of simulation, allowing the study of larger, more detailed models, in less time. PDES is a fine-grained parallel application whose performance and scalability is limited by communication latencies. Traditionally, PDES simulation kernels use message passing; often these simulators are written for distributed environments, and shared memory is used to optimize message passing among processes on the same machine. In this paper, we develop, characterize and optimize a thread-based version of a PDES simulator on three representative multi-core platforms. The multi-threaded implementation eliminates multiple message copying and significantly minimizes synchronization delays. We study the performance of the simulator on three hardware platforms: an Intel Core i7 machine, and a 48-core AMD Opteron Magny-Cours system, and a 64-core Tilera TilePro64. We discover that the three platforms encounter substantially different bottlenecks because of their different architectures. We identify these bottlenecks and propose mechanisms to overcome them. Our results show that multi-threaded implementation improves the performance over an MPI-based version by up to a factor of 3 on the Core i7, 1.4 on the AMD Magny-Cours, and 2.8 on the Tilera Tile64.

LU Factorization with Partial Pivoting for a Multicore System with Accelerators

LU factorization with partial pivoting is a canonical numerical procedure and the main component of the high performance LINPACK benchmark. This paper presents an implementation of the algorithm for a hybrid, shared memory, system with standard CPU cores and GPU accelerators. The difficulty of implementing the algorithm for such a system lies in the disproportion between the computational power of the CPUs, compared to the GPUs, and in the meager bandwidth of the communication link between their memory systems. An additional challenge comes from the complexity of the memory-bound and synchronization-rich nature of the panel factorization component of the block LU algorithm, imposed by the use of partial pivoting. The challenges are tackled with the use of a data layout geared toward complex memory hierarchies, autotuning of GPU kernels, fine-grain parallelization of memory-bound CPU operations and dynamic scheduling of tasks to different devices. Performance in excess of one TeraFLOPS is achieved using four AMD Magny Cours CPUs and four NVIDIA Fermi GPUs.

Approximate weighted matching on emerging manycore and multithreaded architectures

Graph matching is a prototypical combinatorial problem with many applications in high-performance scientific computing. Optimal algorithms for computing matchings are challenging to parallelize. Approximation algorithms are amenable to parallelization and are therefore important to compute matchings for large-scale problems. Approximation algorithms also generate nearly optimal solutions that are sufficient for many applications. In this paper we present multithreaded algorithms for computing half-approximate weighted matching on state-of-the-art multicore (Intel Nehalem and AMD Magny-Cours), manycore (Nvidia Tesla and Nvidia Fermi), and massively multithreaded (Cray XMT) platforms. We provide two implementations: the first uses shared work queues and is suited for all platforms; and the second implementation, based on dataflow principles, exploits special features available on the Cray XMT. Using a carefully chosen dataset that exhibits characteristics from a wide range of applications, we show scalable performance across different platforms. In particular, for one instance of the input, an R-MAT graph (RMAT-G), we show speedups of about [Formula: see text] on [Formula: see text] cores of an AMD Magny-Cours, [Formula: see text] on [Formula: see text] cores of Intel Nehalem, [Formula: see text] on Nvidia Tesla and [Formula: see text] on Nvidia Fermi relative to one core of Intel Nehalem, and [Formula: see text] on [Formula: see text] processors of Cray XMT. We demonstrate strong as well as weak scaling for graphs with up to a billion edges using up to 12,800 threads. We avoid excessive fine-tuning for each platform and retain the basic structure of the algorithm uniformly across platforms. An exception is the dataflow algorithm designed specifically for the Cray XMT. To the best of the authors' knowledge, this is the first such large-scale study of the half-approximate weighted matching problem on multithreaded platforms. Driven by the critical enabling role of combinatorial algorithms such as matching in scientific computing and the emergence of informatics applications, there is a growing demand to support irregular computations on current and future computing platforms. In this context, we evaluate the capability of emerging multithreaded platforms to tolerate latency induced by irregular memory access patterns, and to support fine-grained parallelism via light-weight synchronization mechanisms. By contrasting the architectural features of these platforms against the Cray XMT, which is specifically designed to support irregular memory-intensive applications, we delineate the impact of these choices on performance.

OpenMP task scheduling strategies for multicore NUMA systems

The recent addition of task parallelism to the OpenMP shared memory API allows programmers to express concurrency at a high level of abstraction and places the burden of scheduling parallel execution on the OpenMP run-time system. Efficient scheduling of tasks on modern multi-socket multicore shared memory systems requires careful consideration of an increasingly complex memory hierarchy, including shared caches and non-uniform memory access (NUMA) characteristics. In order to evaluate scheduling strategies, we extended the open source Qthreads threading library to implement different scheduler designs, accepting OpenMP programs through the ROSE compiler. Our comprehensive performance study of diverse OpenMP task-parallel benchmarks compares seven different task-parallel run-time scheduler implementations on an Intel Nehalem multi-socket multicore system: our proposed hierarchical work-stealing scheduler, a per-core work-stealing scheduler, a centralized scheduler, and LIFO and FIFO versions of the Qthreads round-robin scheduler. In addition, we compare our results against the Intel and GNU OpenMP implementations. Our hierarchical scheduling strategy leverages different scheduling methods at different levels of the hierarchy. By allowing one thread to steal work on behalf of all of the threads within a single chip that share a cache, the scheduler limits the number of costly remote steals. For cores on the same chip, a shared LIFO queue allows exploitation of cache locality between sibling tasks as well as between a parent task and its newly created child tasks. In the performance evaluation, our Qthreads hierarchical scheduler is competitive on all benchmarks tested. On five of the seven benchmarks, it demonstrates speedup and absolute performance superior to both the Intel and GNU OpenMP run-time systems. Our run-time also demonstrates similar performance benefits on AMD Magny Cours and SGI Altix systems, enabling several benchmarks to successfully scale to 192 CPUs of an SGI Altix.