Embedding SCI into PVM

The extremely low latencies and high bandwidth results achievable with the Scalable Coherent Interface (SCI) at lowest level encourages its integration into existing Message Passing Environments (MPE's). In combination with Network of Workstations, it can be seen as an alternative for traditional parallel computing with tightly coupled processors with comparable performance at low cost. This paper describes the implementation of PVM using SCI hardware devices and with Linux and WindowsNT as the operating systems. It gives an overview of SCI and the possibilities for PVM to make use of the superior features of SCI compared with the conventional Ethernet. First benchmark results are presented showing the potential performance of this new technology. 1 Parallel Computing over heterogeneous Networks An ongoing trend is to build up network of workstations as an alternative for parallel computing. In this case each node of the parallel computer is an o the-shelf personal computer or server connected via standard network interface cards to a communication network. Compared with the steadily increasing computing power of processors, network technology could not cope with processing performance. However increasing number and size of message exchange of ne grain parallel algorithms worsen the situation. SCI (Scalable Coherent Interface) is a new communication technology which allows the e ective coupling of modular components as well as to cluster large and distributed computer systems. The SCI protocol supports message passing as well as distributed shared memory access. High-speed unidirectional pointto-point links can achieve a communication bandwidth of up to 1 GByte/s. A more important factor for small messages is the low latency of only some micro seconds. Combined with the computing-power of today available low-cost workstations, these superior properties allow the construction of cost-e ective high-speed multiprocessor clusters that can achieve supercomputer performance. According to the promising basic performance properties of SCI we are attempting to make use of these features by integration into a Message Passing Environment. PVM has been chosen to be expanded to SCI because it provides both Linux and WindowsNT compatibility. The goal of the port is to develop a SCI-library which has to be linked instead of the currently used socket library. But also the possibility to communicate with non-SCI workstations should be kept open. This extensibility has to be transparent to the application and to be controlled by the PVM library. Since PVM supports shared memory principles as well as the exchange of messages over sockets the later mentioned features of PVM are stated according to the UNIX avored workstations. In the following we present results which are based on a SCI workstation cluster which consists of several symmetrical multiprocessor systems (SMP), double processor as well as standard high volume servers with four processors based on the INTEL PentiumPro architecture, with a total of 20 processors. Both operating systems Linux and WindowsNT are supported. 2 SCI Communication Technology SCI (Scalable Coherent Interface) is the international standard IEEE #15961992 for computer-bus-like services on a ring-based network [IEE93]. It supports distributed shared memory (DSM) and message passing for loosely and tighly coupled systems. Optionally cache coherent transactions are supported to implement CC-NUMA (Cache Coherent Non-Uniform Memory Access) architectures. Such cache coherent SCI interfaces are implemented in Sequent's NUMA-Q multiprocessr systems with the IQ-Link architecture and in the HP-Convex SPPsystems with the CTI network [Seq97, Con96]. Non cache coherent SCI communications cards for the standard computer I/O busses PCI, SBUS, and VME are implemented by Dolphin Interconnect Solutions, CERN, and Technische Universitat M unchen [Dol96, AHKL96]. The SCI IEEE standard allows besides simple ringlets also more complex network topologies. Dolphins provides a four port switch which connects four separate SCI ringlets to a single network [Dol97].