Writing parallel applications for computational grids is a challenging task. To achieve good performance, algorithms designed for local area networks must be adapted to the differences in link speeds. An important class of algorithms are collective operations, such as broadcast and reduce. We have developed MAGPIE, a library of collective communication operations optimized for wide area systems. MAGPIE's algorithms send the minimal amount of data over the slow wide area links, and only incur a single wide area latency. Using our system, existing MPI applications can be run unmodified on geographically distributed systems. On moderate cluster sizes, using a wide area latency of 10 milliseconds and a bandwidth of 1 MByte/s, MAGPIE executes operations up to 10 times faster than MPICH, a widely used MPI implementation; application kernels improve by up to a factor of 4. Due to the structure of our algorithms, MAGPIE's advantage increases for higher wide area latencies.

This work provides a systematic study of the impact of communication performance on parallel applications in a high performance network of workstations. We develop an experimental system in which the communication latency, overhead, and bandwidth can be independently varied to observe the effects on a wide range of applications. Our results indicate that current efforts to improve cluster communication performance to that of tightly integrated parallel machines results in significantly improved application performance. We show that applications demonstrate strong sensitivity to overhead, slowing down by a factor of 60 on 32 processors when overhead is increased from 3 to 103 s. Applications in this study are also sensitive to per-message bandwidth, but are surprisingly tolerant of increased latency and lower per-byte bandwidth. Finally, most applications demonstrate a highly linear dependence to both overhead and per-message bandwidth, indicating that further improvements in communication performance will continue to improve application performance. 1

Clusters of multiprocessors, or Clumps, promise to be the supercomputers of the future, but obtaining high performance on these architectures requires an understanding of interactions between the multiple levels of interconnection. In this paper,wepresent the rst multi-protocol implementation of a lightweight message layer|a version of Active Messages-II running on a cluster of Sun Enterprise 5000 servers connected with Myrinet. This research brings together several pieces of high-performance interconnection technology: bus backplanes for symmetric multiprocessors, low-latency networks for connections between machines, and simple, user-level primitives for communication. The paper describes the shared memory message-passing protocol and analyzes the multiprotocol implementation with both microbenchmarks and Split-C applications. Three aspects of the communication layer are critical to performance: the overhead of cache-coherence mechanisms, the method of managing concurrent access, and the cost of accessing state with the slower protocol. Through the use of an adaptive polling strategy, the multi-protocol implementation limits performance interactions between the protocols, delivering up to 160 MB/s of bandwidth with 3.6 microsecond end-to-end latency. Applications within an SMP bene t from this fast communication, running up to 75 % faster than on a network of uniprocessor workstations. Applications running on the entire Clump are limited by the balance of NIC's to processors in our system, and are typically slower than on the NOW. These results illustrate several potential pitfalls for the Clumps architecture. 1

With the use of state and memory reduction techniques in verification by explicit state enumeration, runtime becomes a major limiting factor. We describe a parallel version of the explicit state enumeration verifier Murφ for distributed memory multiprocessors and networks of workstations that is based on the message passing paradigm. In experiments with three complex cache coherence protocols, parallel Murφ shows close to linear speedups, which are largely insensitive to communication latency and bandwidth. There is some slowdown with increasing communication overhead, for which a simple yet relatively accurate approximation formula is given. Techniques to reduce overhead and required bandwidth and to allow heterogeneity and dynamically changing load in the parallel machine are discussed, which we expect will allow good speedups when using conventional networks of workstations.

The POEMS project is creating an environment for end-to-end performance modeling of complex parallel and distributed systems, spanning the domains of application software, runtime and operating system software, and hardware architecture. Towards this end, the POEMS framework supports composition of component models from these different domains into an end-to-end system model. This composition can be specified using a generalized graph model of a parallel system, together with interface specifications that carry information about component behaviors and evaluation methods. The POEMS Specification Language compiler, under development, will generate an end-to-end system model automatically from such a specification. The components of the target system may be modeled using different modeling paradigms (analysis, simulation, or direct measurement) and may be modeled at various levels of detail. As a result, evaluation of a POEMS end-to-end system model may require a variety of eval...

In this thesis, we formalize the concept of an implicitly-controlled system, also referred to as an implicit system. In an implicit system, cooperating components do not explicitly contact other components for control or state information; instead, components infer remote state by observing naturally-occurring local events and their corresponding implicit information, i.e., information available outside of a defined interface. Many systems, particularly in distributed and networked environments, have leveraged implicit control to simplify the implementation of services with autonomous components. To concretely demonstrate the advantages of implicit control, we propose and implement implicit coscheduling, an algorithm for dynamically coordinating the time...

Parallel algorithm designers need computational models that take first order system costs into account, but are also simple enough to use in practice. This paper introduces the LoPC model, which is inspired by the LogP model but accounts for contention for message processing resources in parallel algorithms on a multiprocessor or network of workstations. LoPC takes the , and parameters directly from the LogP model and uses them to predict the cost of contention, .

There has been a great deal of interest recently in the development of general-purpose bridging models for parallel computation. Models such as the BSP and LogP have been proposed as more realistic alternatives to the widely used PRAM model. The BSP and LogP models imply a rather different style for designing algorithms when compared with the PRAM model. Indeed, while many consider data parallelism as a convenient style, and the shared-memory abstraction as an easyto-use platform, the bandwidth limitations of current machines have diverted much attention to message-passing and distributed-memory models (such as the BSP and LogP) that account more properly for these limitations. In this paper we consider the question of whether a shared-memory model can serve as an effective bridging model for parallel computation. In particular, can a shared-memory model be as effective as, say, the BSP? As a candidate for a bridging model, we introduce the Queuing Shared-Memory (QSM) model, which accounts for limited communication bandwidth while still providing a simple shared-memory abstraction. We substantiate the ability of the QSM to serve as a bridging model by providing a simple work-preserving emulation of the QSM on both the BSP, and on a related model, the (d, x)-BSP. We present evidence that the features of the QSM are essential to its effectiveness as a bridging model. In addition, we describe scenarios

We present MRNet, a software-based multicast/reduction network for building scalable performance and system administration tools. MRNet supports multiple simultaneous, asynchronous collective communication operations.

Previous studies of application usage show that the performance of collective communica-tions are critical for high performance computing and are often overlooked when compared to the point-to-point performance. In this paper we attempt to analyze and improve collective communication in the context of the widely deployed MPI programming paradigm by extending accepted models of point-to-point communication, such as Hockney, LogP/LogGP, and PLogP. The predictions from the models were compared to the experimentally gathered data and our findings were used to optimize the implementation of collective operations in the FT-MPI library. 1