Many scientific codes consist of memory bandwidth bound kernels — the dominating factor of the runtime is the speed at which data can be loaded from memory into the Arithmetic Logic Units. Generally Programmable Graphics Processing Units (GPGPUs) and other accelerator devices such as the Intel Xeon Phi offer an increased memory bandwidth over CPU architectures. However, as with CPUs, the peak memory bandwidth is often unachievable in practice and so benchmarks are required to measure a practical upper bound on expected performance. We present GPU-STREAM as an auxiliary tool to the standard STREAM benchmark to provide cross-platform comparable results of achievable memory bandwidth between multiand many-core devices. I. MEASURING MEMORY BANDWIDTH The STREAM Benchmark [1] measures the time taken for each of four simple kernels to be run (α is a scalar constant): 1) Copy: c[i] = a[i] 2) Multiply: b[i] = αc[i] 3) Add: c[i] = a[i] + b[i] 4) Triad: a[i] = b[i] + αc[i] It is simple to calculate how many bytes each of these kernels requires to be read from and written to memory, assuming perfect caching. Let β be the number of bytes to represent one element — for double precision β = 8. For an array of length N , copy and multiply move 2Nβ bytes, and add and triad move 3Nβ bytes. The ordering of the kernels ensures caches are invalidated, and as long as the arrays are large enough the data must be reloaded from main memory. The achieved sustained memory bandwidth can be found by dividing these numbers by the time to execute the corresponding kernel. GPU-STREAM implements these kernels in both the OpenCL and CUDA programming frameworks. This allows the benchmark to be used across a wide range of hardware from a wide range of vendors.
We present a new low-latency system area network that provides the ultra-high bandwidth needed to fuse a collection of large SMP servers into a capability cluster. The network adapter exports a remote shared memory (RSM) model that supports low latency kernel bypass messaging. The Sun™ MPI library uses the RSM interface to implement a highly efficient memory-to-memory messaging protocol in which the library directly manages buffers and data structures in remote memory. This allows flexible allocation of buffer space to active connections, while avoiding resource contention that could otherwise increase latencies. We discuss the characteristics of the interconnect, describe the MPI protocols, and measure the performance of a number of MPI benchmarks. Our results include MPI inter-node bandwidths of almost 3 Gigabytes per second and MPI ping-pong latencies as low as 3.7 microseconds.
Although workstation clusters are a common platform for high-performance computing (HPC), they remain more difficult to manage than sequential systems or even symmetric multiprocessors. Furthermore, as cluster sizes increase, the quality of the resource-management subsystem — essentially, all of the code that runs on a cluster other than the applications — increasingly impacts application efficiency. In this paper, we present STORM, a resource-management framework designed for scalability and performance. The key innovation behind STORM is a software architecture that enables resource management to exploit low-level network features. As a result of this HPC-application-like design, STORM is orders of magnitude faster than the best reported results in the literature on two sample resource-management functions: job launching and process scheduling.
We present Salinas, a scalable implicit software application for the finite element static and dynamic analysis of complex structural real-world systems. This relatively complete engineering software with more than 100,000 lines of C++ code and a long list of users sustains 292.5 Gflop/s on 2,940 ASCI Red processors, and 1.16 Tflop/s on 3,375 ASCI White processors.
In wide area computing systems, it is often desirable to create remote read-only copies (replicas) of files. Replication can be used to reduce access latency, improve data locality, and/or increase robustness, scalability and performance for distributed applications. We define a replica location service (RLS) as a system that maintains and provides access to information about the physical locations of copies. An RLS typically functions as one component of a data grid architecture. This paper makes the following contributions. First, we characterize RLS requirements. Next, we describe a parameterized architectural framework, which we name Giggle (for GIGa-scale Global Location Engine), within which a wide range of RLSs can be defined. We define several concrete instantiations of this framework with different performance characteristics. Finally, we present initial performance results for an RLS prototype, demonstrating that RLS systems can be constructed that meet performance goals.
A major challenge in Internet edge router design is to support both high packet forwarding performance and versatile and efficient packet processing capabilities. The thesis of this research project is that a cluster of PCs connected by a high speed system area network provides an effective hardware platform for building routers to be used at the edges of the Internet. This paper describes a scalable and extensible edge router architecture called Panama, which supports a novel aggregate route caching scheme, a real-time link scheduling algorithm whose performance overhead is independent of the number of real-time flows, a highly efficient kernel extension mechanism to safely load networking software extensions dynamically, and an integrated resource scheduler which ensures that real-time flows with additional packet processing requirements still meet their end-to-end performance requirements. This paper describes the implementation and evaluation of the first Panama prototype based on a cluster of PCs and Myrinet.
As an entry for the 2002 Gordon Bell performance prize, we report the performance achieved on the GRAPE-6 system for a simulation of the early evolution of the protoplanet-planetesimal system of the Uranus-Neptune region. GRAPE-6 is a special-purpose computer for astrophysical N-body calculations. The present configuration has 2048 custom pipeline chips, each containing six pipeline processors for the calculation of gravitational interactions between particles. Its theoretical peak performance is 63.4 Tflops. The actual performance obtained was 29.5 Tflops, for a simulation of the early evolution of outer Solar system with 1.8 million planetesimals and two massive protoplanets.
In this paper, we present the Active Harmony automated runtime tuning system. We describe the interface used by programs to make applications tunable. We present the Library Specification Layer which helps program library developers expose multiple variations of the same API using different algorithms.The Library Specification Language helps to select the most appropriate program library to tune the overall performance. We also present the optimization algorithm used to adjust parameters in the application and the libraries. Finally, we present results that show how the system is able to tune several real applications. The automated tuning system is able to tune the application parameers to within a few percent of the best value after evaluating only 11 out of over 1,700 possible configurations.
The SmartPointer system provides a paradigm for utilizing multiple light-weight client endpoints in a real-time scientific visualization infrastructure. Together, the client and server infrastructure form a new type of data portal for scientific computing. The clients can be used to personalize data for the needs of the individual scientist. This personalization of a shared dataset is designed to allow multiple scientists, each with their laptops or iPaqs to explore the dataset from different angles and with different personalized filters. As an example, iPaq clients can display 2D derived data functions which can be used to dynamically update and annotate the shared data space, which might be visualized separately on a large immersive display such as a CAVE. Measurements are presented for such a system, built upon the ECho middleware system developed at Georgia Tech.
The model for data mining on streaming data assumes that there is a buffer of fixed length and a data stream of infinite length and the challenge is to extract patterns, changes, anomalies, and statistically significant structures by examining the data one time and storing records and derived attributes of length less than N . As data grids, data webs, and semantic webs become more common, mining distributed streaming data will become more and more important. The first step when presented with two or more distributed streams is to merge them using a common key. In this paper, we present two algorithms for merging streaming data using a common key. We also present experimental studies showing these algorithms scale in practice to OC-12 networks.
In this paper, we develop a parallel Fast Multipole Method (FMM) based solution for computing the scattered electromagnetic fields from a Perfect Electrically Conducting (PEC) surface. The main contributions of this work are the development of parallel algorithms with the following characteristics: 1) provably efficient worst-case run-time irrespective of the shape of the scatterer, 2) communication-efficiency, and 3) guaranteed load balancing within a small constant factor. We have developed a scalable, parallel code and validated it against surfaces for which solution can be computed analytically, and against serial software. The efficiency and scalability of the code is demonstrated with experimental results on an IBM xSeries cluster. Though developed in the context of this particular application, our algorithms can be used in other applications involving parallel FMM.
We investigate remapping multi-dimensional arrays on cluster of SMP architectures under OpenMP, MPI, and hybrid paradigms. Traditional method of array transpose needs an auxiliary array of the same size and a copy back stage. We recently developed an in-place method using vacancy tracking cycles. The vacancy tracking algorithm outperforms the traditional 2-array method as demonstrated by extensive comparisons. The independence of vacancy tracking cycles allows efficient parallelization of the in-place method on SMP architectures at node level. Performance of multi-threaded parallelism using OpenMP are tested with different scheduling methods and different number of threads. The vacancy tracking method is parallelized using several parallel paradigms. At node level, pure OpenMP outperforms pure MPI by a factor of 2.76. Across entire cluster of SMP nodes, the hybrid MPI/OpenMP implementation outperforms pure MPI by a factor of 4.44, demonstrating the validity of the parallel paradigm of mixing MPI with OpenMP.
We propose data sharing facility for data intensive scientific research, "Data Reservoir"; which is optimized to transfer huge amount of data files between distant places fully Utilizing multi-gigabit backbone network. In addition, "Data Reservoir" can be used as an ordinary UNIX server in local network without any modification of server softwares. We use low-level protocol and hierarchical striping to realize (1) separation of bulk data transfer and local accesses by cashing, (2) file-system transparency, I.e. interoperable whatever in higher layer than disk driver, including file system. (3) scalability for network and storage. This paper shows our design, implementation using iSCSI protocol [1] and their performances for both 1Gbps model in the real network and 10Gbps model in our laboratory.
High performance computing, storage, visualization, and database infrastructures are increasing geometrically in complexity as scientists move towards grid-based computing. While this is natural, it has the effect of pushing computational capabilities beyond the reach of scientists because of the time needed to harness the infrastructure. Hiding the complexity of networked resources becomes essential if scientists are to utilize them effectively. In this work, we describe our efforts to integrate various computational chemistry components into a scientific computing environment. We briefly describe improvements we have made to individual components of the chemistry environment as well as future directions, followed by a more in-depth discussion of our strategy for integration into a grid workflow environment based on web services, which enables access to remote resources while shielding users from the complexities of the grid infrastructures. A preliminary schema for storing data obtained from computational chemistry calculations is also described.
