Monthly Archives: March 2006

To reconcile the parallelism in both tasks (OpenMP) and data (HPF), a different programming model has emerged that claims to have the best of both. It is the partitioned global address space model, and has been applied a variety of languages, the most widely used being Unified Parallel C (UPC).

In UPC, by default, all variables are private and every instruction is executed by each thread. However, the user may declare that some data be …

A grid is a wide-area network of computers used to process parts of the same problem. Unlike a cluster, grids tend to be loosely coupled in that there usually is no central control. The notion of grid computing essentially came from the observation that most PCs are underused, which means that their computing powers could be shared and combined to form a supercomputer on-the-fly.

The technology enabler for grid computing tends to come from open …

In the task-parallel model represented by OpenMP, the user specifies the distribution of iterations among processors and then the data travels to the computations. In data-parallel programming, the user specifies the distribution of arrays among processors, and then only those processors owning the data will perform the computation. In OpenMP’s master / slave approach, all code is executed sequentially on one processor by default. In data-parallel programming, all code is executed on every processor in …

The term “personal supercomputer” is a marketing ploy that has been applied to a number of higher performance systems. Apple, for example, advertised its PowerMac G4 (a uniprocessor, no less) in this fashion. However, there are some more legitimate uses of the term. Tyan recently announced its own personal supercomputer under the Typhoon banner; that product contains up to eight dual-core CPUs. These devices are small enough to fit on a desk and are …

Virtualization is a technique that allows multiple operating systems to run simultaneously on the same hardware. Consider, for example, a fault-tolerant server in which several instances of Linux are executing. If a fault happens to wipe out one of those instances, the rest keep running unabated. In this regard, virtualization is to operating systems what multiprocess environments are to user applications.

One approach to achieve this is by emulating ALL of the underlying hardware. Such is …

Supercomputers traditionally have many processors linked together with the goal that multiple CPUs can perform multiple actions simultaneously, thereby greatly speeding-up execution. An added benefit is that multiple CPUs allow for redundancy, which increases the availability (up-time) of the system. Specially built systems from Cray and others are especially effective, but are very expensive because they employ a number of proprietary components.

Clusters, unlike traditional supercomputers, employ only commodity components. They are simply regular PCs …

The computing industry has grown at an exceptional pace over the past four decades. Moore’s Law about doubling computer speeds every eighteen months will probably hold for another decade or two. After that, transistors will be about the size of atoms and thus will not shrink anymore. This is the first roadblock to continuing performance improvements indefinitely. The solution here will have to come from an alternative computing source, such as quantum mechanics. Such a …

Most high-performance compilers aim to parallelize loops to speed-up technical codes. Automatic parallelization is possible but extremely difficult because the semantics of the sequential program may change. Therefore, most users provide some clues to the compiler. A very common method is to use a standard set of directives known as OpenMP, in which the user expresses which sections of the code are to be parallel via a pragma.

Ordinarily with OpenMP, the software is executed …

The two major classes of users are technical customers and enterprise customers. The former comprises scientists and engineers who need to perform number crunching; examples include climate prediction, protein folding simulations, oil and gas discovery, defense and aerospace work, automotive design, financial forecasting, etc. The later category encompasses the corporate data center that stores customer records, inventory management, and employee details.

Technical customers usually require the Message Passing Interface (MPI), whereas enterprise customers require an …

Traditionally, computing devices are connected together in a bus, the most popular standard of which will soon be PCI Express. A bus-based architecture tends to have poor latency, though for most applications this is acceptable. Sometimes, however, when the application has more processes than data, numerous small messages tend to be a major factor in performance, and thus latency becomes critical. For these applications, PCI Express may be a weight.

Indeed, one of the original goals …