Towards a Tier-3 testbench configuration

At present Scalable Computing Clusters (CC), ranging from a cluster of (homogeneous) servers to a Network of (heterogeneous) Workstations (NOW), are rapidly becoming the standard platforms for executing demanding applications, high performance and interactive computing. The main attractiveness of such systems is that they are made of affordable, low-cost, commodity hardware, e.g. Pentium based Personal Computers (PCs), and fast LANs e.g. Fast Ethernet. Other advantages are that these systems are scalable, i.e., can be tuned to available budget and needs, and that they use standard software components, i.e., UNIX, PVM and the MPI parallel programming environments. Unlike MPPs (Massive Parallel Processors), which allow a single user per partition, CCs are geared for multi-user, time-sharing environment. In order to make CC systems as easy to program, manage, and use as an SMP (Sy metric Multi Processor), it is necessary to develop means for global (cluster-wide) resource allocation and sharing that can respond to resource availability, distribute the workload dynamically and utilize the available, cluster-wide resources efficiently and transparently. The main techniques for dynamic work distribution are load-balancing and load-sharing, that use some form of remote execution and process migration. These techniques can perform initial distribution of processes among the nodes and redistribution of the work load when the system becomes unbalanced.Taking into consideration our requirements, both hardware and software, necessary for successfully acquitting, managing, and analysing large amount of experimental data, our hardware solution is based on the ANSI/TIA/EIA-568-A structured cabling standard with a star-wired topology, using switched Fast Ethernet technology. The main communication equipment is an Allied Telesyn FS708 Fast Ethernet switch which connects our computers with a full duplex 100 Mbps bandwidth. A schematic view of our group network is presented. As a next step we shall implement the MOSIX Cluster software on our Linux computers, and test the network performance (real maximum bandwidth, latency, etc.). (authors)