Brian Vinter

Java PastSet: a structured distributed shared memory system

The architecture and performance of a Java implementation of a structured distributed shared memory system, PastSet, is described. The PastSet abstraction allows programmers to write applications that run efficiently on different architectures, from clusters to widely distributed systems. PastSet is a tuple-based three-dimensional structured distributed shared memory system, which provides the programmer with operations to perform causally ordered reads and writes of tuples to a virtual structured memory space called the PastSet. It is shown that the original, native code, PastSet is able to outperform MPI and PVM when running real applications and that the design translates into Java so that Java PastSet is a qualified competitor to other cluster application programming interfaces for Java.

Minimum intrusion grid - the simple model

This paper describes the key ideas in the grid model: MiG (minimum intrusion grid). The simple model is a major milestone in developing the full MiG concept and when completely implemented the simple model includes a full featured grid solution. Besides describing the general MiG idea, the design of the simple model is specified as well as the main considerations behind the implementation.

Transparent remote file access in the minimum intrusion grid

This paper describes the implementation of a thin user-level layer to be installed on grid resources. The layer fits in the minimum intrusion grid design by imposing as few requirements on the resource as possible and communicates with the server using only trusted and widely used protocols. The model offers transparent, on-demand remote file access. By catching all application operations on input files, these operations are directed towards the remote copy on the server, thus eliminating the need for transferring the complete input file. This implementation is targeted at the minimum intrusion grid project, which strives for minimum intrusion on the resource executing a job. In minimum intrusion lies that a client need not install any dedicated grid software, forcing the proposed model to use a user-level layer that automatically overrides GLIBG I/O calls.

The grid block device: performance in LAN and WAN environments

We present initial results from the implementation of the Grid Block Device, a distributed, replicated block device for the Grid based on the the replication algorithm of Y. Amir[1]. It supports application-specific replication strategies and consistency models. Although the presented WAN results are dissapointing because of issues with the underlying group communication library, we believe the LAN results show that the Grid Block Device could become a viable solution for truly distributed storage.

Direct application access to Grid storage

This paper describes the ideas behind and the implementation of a thin user‐level layer to be installed on Grid resources. The layer fits in the minimum intrusion Grid design by imposing as few requirements on the resource as possible and communicates with the server using only trusted and widely used protocols. The model offers transparent, on‐demand remote file access. By catching all application operations on files, these operations are directed towards the remote copy on the server, thus eliminating the need for transferring the complete file. This implementation is targeted at the minimum intrusion Grid project, which strives for minimum intrusion on the resource executing a job. ‘Minimum intrusion’ implies that a client need not install any dedicated Grid software. Hence, the proposed model is forced to use a user‐level layer that automatically overrides the native I/O calls. Copyright

Cluster computing as a teaching tool

Publisher Summary This chapter discusses cluster computing. Clusters are sets of loosely coupled computers administered and employed in parallel to solve shared tasks. Clusters of common off-the-shelves (COTS) components constitute a class of cost-effective parallel computers, allowing maintained harvesting of advances in processor and system technologies. The chapter describes how students in operating systems courses as well as advanced courses on parallel computing may use a network of workstations and MPI to learn about synchronization and parallelization issues. It then explains how an advanced course on distributed and parallel systems use clusters to support cheap robots with compute cycles. The robots used are made by Lego, and the system used demonstrates the principles and practice of distributed computing in general, and high performance parallel computing in particular. Software used for the projects in this course includes JAVA-MPI, JAVA RMI, RPC, Tspacesry, and Linda.

Embarrassingly Parallel Applications on a Java Cluster

This paper proposes a new benchmark suite for use with research in Java for HPC, and introduces the first set of applications that are all of the embarrassingly parallel type. The applications are implemented using threads and four kinds of networked IPC mechanisms. The tested APIs include both native Java NIPCs and interfaced to native code NIPCs. The results show that there is huge differences in the scalability of the tested APIs, even for embarrassingly parallel applications.

Auto-tuning for large-scale image processing by dynamic analysis method on multicore platforms

This paper describes a general-purpose method of improving execution performance of the in-memory data, particularly in the case of large-scale image processing on different multicore platforms. To process large-scale arrays, the method of tiling is widely used to achieve high performance. However, frequently accessing the memory system by multithreads is bound to cause system bottleneck. Our optimisation strategies are automatic thread scheduling and data/task partitioning. Those methods that attempt to take advantage of spatial and temporal locality can reduce memory traffic remarkably. According to the hardware configurations, a scheduler automatically partitions the images into tiled blocks of pre-determined size. Then it fuses all the operations for the same blocks to reduce the rate of cache miss. The parallel task execution is more effective than other traditional parallel libraries, such as openMP. Moreover, the optimisation on space-filling curves that optimises the locality of neighbouring tiled blocks can also contribute to the fast memory access.

Performance and Portability of the SciBy Virtual Machine

The Scientific Bytecode Virtual Machine is a virtual machine designed specifically for performance, security, and portability of scientific applications deployed in a Grid environment. The performance overhead normally incurred by virtual machines is mitigated using native optimized scientific libraries, security is obtained by sandboxing techniques. Lastly, by executing platform-independent bytecodes, the machine is highly portable. To evaluate the machine, we demonstrate several use-case scenarios from some of the intended application domains. Further, we show the ease of porting the machine and distributing its jobs to a variety of predominant architectures and compare the results with native execution.

The one-click grid-resource model

This paper introduces the One-Click Grid resource, which allows any computer with a Java enabled web browser to safely provide resources to Grid without any software installation. This represents a vast increase of the number of potential Grid resources that may be made available to help public interest research. While the model does make restrictions towards the application writer, the technology provides a real Grid model and supports arbitrary binaries, remote file access and semitransparent checkpointing. Performance numbers show that the model is usable even with browsers that are connected to the Internet through relatively weak links, i.e. 512 kb/s upload speeds. The resulting system is in use today, and freely available to any research project.