Despite technological advantages in microprocessors and network technology over the last few years, commercially-available networks of workstations (NOWs) contain inherent communication bottlenecks. Traditional layered network protocols will inevitably fail to achieve high throughput if they access data several times. As a result, applications on NOWs often fail to observe the performance speed-up that might be expected. Network protocols which avoid routing through the kernel can remove this limit on communication performance and support very high transmission speeds, turning NOWs into an attractive alternative to Massively-Parallel Processors. This paper investigates the improvements in performance of an entire NOW system that are possible by changing communication protocols. A commodity workstation cluster with a zero-copy communication protocol (Basic Interface Parallelism) is examined and compared with both conventional protocol implementations and parallel systems in order to un...

Wormhole-routing high-speed LANs (e.g. Myrinet), increasingly used to build high-performance Network of Workstations, do not usually provide hardware support to multicast communication, that must instead be implemented in software by sending several point-to-point messages. Network Interface protocols allow an efficient software implementation of multicast communication, but require flow control techniques (to avoid buffer overflows) that may lead to a severe performance degradation. In this paper we address the problem of designing reliable and efficient Network Interface multicast protocols. We propose a flow control technique that provides the basis for 4 multicast protocols, and we perform extensive simulations to characterize the impact that various factors (such as network topology, presence of unicast traffic, composition of multicast groups, etc.) have on their performance. Interestingly, our simulation experiments sometimes lead to observations that are in contrast wi...

In this paper, we will introduce two different physical structures and topologies to realize an ethernet layer-2 switch. To evaluate these two structures and produce throughput projections we have utilized MPI. This simulation environment has helped us identify performance bottlenecks and to give an overall estimation of the switch throughput. Our results show that for a hardware implementation of this switch in .18um technology, assuming a 200 MHz clock rate for the processing elements and 100 MHz clock rate for the memory blocks, our design can support multiple gigabit channel ports.

2002 ”Communication bandwidth is like candy given to a small child – you give the child one piece of candy and he wants a pound.” Clusters of Personal Computers (CoPs) offer the best compute performance at the lowest price. Workstations with ’Gigabit networking to the Desktop ’ can enable a new game of multimedia applications that benefit from higher communication bandwidth and lower latency. In order to reach the full Gigabit/s speed on normal PCs with their typically weak memory subsystems it requires either additional hardware for protocol processing or alternatively, a highly efficient software system that circumvents data copies. In this dissertation we successfully introduced speculation techniques into system software design and managed to implement a clean zero-copy solution entirely in software that runs with commodity network interface cards (NICs) like the ubiquitous and cheap Gigabit Ethernet adapters, using the standard TCP/IP protocol and the socket API. The implementation techniques are similar to the ones that are already widely used in