The U-Net communication architecture provides processes with a virtual view of a network interface to enable userlevel access to high-speed communication devices. The architecture, implemented on standard workstations using offthe-shelf ATM communication hardware, removes the kernel from the communication path, while still providing full protection. The model presented by U-Net allows for the construction of protocols at user level whose performance is only limited by the capabilities of network. The architecture is extremely flexible in the sense that traditional protocols like TCP and UDP, as well as novel abstractions like Active Messages can be implemented efficiently. A U-Net prototype on an 8-node ATM cluster of standard workstations offers 65 microseconds round-trip latency and 15 Mbytes/sec bandwidth. It achieves TCP performance at maximum network bandwidth and demonstrates performance equivalent to Meiko CS-2 and TMC CM-5 supercomputers on a set of Split-C benchmarks. 1

Fast Ethernet and ATM are two attractive network technologies for interconnecting workstation clusters for parallel and distributed computing. This paper compares network interfaces with and without programmable co-processors for the two types of networks using the U-Net communication architecture to provide low-latency and high-bandwidth communication. U-Net provides protected, user-level access to the network interface and offers application-level round-trip latencies as low as 60sec over Fast Ethernet and 90sec over ATM. The design of the network interface and the underlying network fabric have a large bearing on the U-Net design and performance. Network interfaces with programmable co-processors can transfer data directly to and from user space while others require aid from the operating system kernel. The paper provides detailed performance analysis of U-Net for Fast Ethernet and ATM, including application-level performance on a set of Split-C parallel benchmarks. These results s...

A Network of Workstations (NOW) has become an important distributed platform for large-scale scientific computations. A practical NOW system is heterogeneous and nondedicated, where computing power varies among the workstations and multiple jobs may interact with each other during execution. In this paper, we present the design and implementation of a simulation system for a nondedicated heterogeneous NOW. This simulator provides many options to users to specify and quantify system architectures, network heterogeneity and time-sharing factors, such as speeds of different processors, memory organizations, network topology, communication structures, and workload distributions. The simulator also supports execution of message-passing parallel programs written in C and the PVM library. The software structure of the simulator is well-modularized and highly extensible, which makes it easy to integrate other existing processor, memory and network simulators. 1 Introduction A Network of Works...

this paper, every packet is broken into a number of flits, and buffering, forwarding and flowcontrol are performed at the flit level. The flits of a packet are sent consecutively over a channel, so the flits of two packets are never interleaved. Flits themselves are actually transmitted a phit (physical transfer unit) at a time, which is typically the size of the link width, something that can be transferred in a single clock cycle. Figure 2 illustrates how a message is partitioned into packets, into flits and then into phits. Network routers connect the processing nodes to the network and manage the links to the neighboring nodes. A router normally contains communication processing logic as well as a set of buffers to hold flits. It handles all communication related tasks to allow computation (by the processor) and communication at the node to take place concurrently. These communication tasks include relaying packets from one node to the next in the direction of the packets' destination node(s) (switching and routing), preventing buffer overflow (flow-control), removing packets from the network if destined for the local node, and injecting packets from the local node Direct Multiprocessor Networks \Delta 3 H D1 D2 D3 H H H F1 F2 P1 P2 Packets Flits Phits Message Message Routing unit Switching and flow-control unit Transmission unit Application unit P1 P2 P1 Fig. 2. The figure illustrates the message, packets, flits and phits in direct networks. into the network (switching). In addition, some routers also assemble packets into messages and disassemble messages into packets. The behavior of a direct network is determined primarily by how it does switching, routing and flow-control. Switching is the mechanism by which a router removes a packet from its input link and p...

Popularized by Napster and Gnutella file sharing solutions, peer to peer (P2P) computing has suddenly emerged at the forefront of Internet computing. The basic notion of cooperative computing and resource sharing has been around for quite some time, although these new applications have opened up possibilities of very flexible web-based information sharing. This paper provides a framework for classifying current and future P2P technologies and discusses the unresolved issues in each case. The main motivation for the taxonomy is to identify basic characteristics of peer to peer applications so that the infrastructure to support peer to peer computing can concentrate on these basic characteristics. The paper also presents some directions for future research in this area.

I-Cluster is a research initiative from HP Labs Grenoble in partnership with INRIA (ID-IMAG Laboratory) . It provides a distributed Peer-to-Peer framework of tools that transparently take advantage of unused network resources and federate them to crystallize into specific virtual functions. These functions are compute-inten sive services, which are statically decomposed to fit the ICluster framework, then dynamically recomposed upon users needs. Typical I-Cluster functions are supercomputing, content rendering, content distribution. The I-Cluster experimentation platform, based on 225 typical desktop machines has been benchmarked as the 385th most powerful supercomputer in the world using Linpack, making it the first cluster based on mainstream technologies to enter the TOP500 in May 2001 .