providing access via the system area network to remote storage resources, e.g., an NFS server on one of the nodes. This may be a practical solution for small Beowulf systems, but as the system grows, it is easy to overwhelm a single server.
The results of computational activities performed on a Beowulf node must be presented to the node's external environment during and after a computation. This requires communication with peripheral devices like video monitors, printers and external networks. Furthermore, users need access to the system to start jobs and to monitor and control jobs in progress. System managers may need console access, the ability to install software distributions on CD-ROM or other media, or backup data to tape or other archival storage. The requirements are served by the I/O subsystem of the node. On today's PCs, these devices usually share the PCI bus, with some low-performance devices using the older ISA bus.
In a Beowulf system it is typical that only one or two nodes have extensive I/O capabilities beyond communication on the system area network. All external interaction is then funneled through these worldly nodes. The specific I/O requirements vary greatly from installation to installation, so a precise specification of the peripherals attached to a worldly node is impossible. We can, however, make firm recommendations about the I/O requirements of internal or compute nodes. The majority of nodes in a Beowulf system lack the personality of a worldly node. They have one major I/O requirement, which is to communicate with one another. The hardware and software involved in interprocessor communication is the subject of Chapter 5. For now, we will simply observe that the processor communicates with the network through the network interface controller (NIC) attached to a high-speed bus.
3.1.1 Principal Specifications
In selecting the proper node configuration for a new Beowulf, the choices can appear overwhelming. Fortunately, there is a small number of critical parameters that largely characterize a particular Beowulf node. These parameters usually relate to a few peak capabilities or capacities and are only roughly predictive of the performance of any given application or workload. Nonetheless, they are widely used and provide a reasonable calibration of the price/performance tradeoff space.
Processor clock rate The frequency (MHz) of the primary signal within the processor that determines the rate at which instructions are issued.
Peak floating point performance The combination of the clock rate and the number of floating point operations that can be issued and/or retired per instruction (Mflops).
