With the appearence of INTERNET, all the computer servers on the INTERNET fors a super-parallel computer system, from the global point of view. And each day, there will be hundreds of computers added into the INTERNET. Millions of computers provide a very strong computing ability. But the problem is how to get use this global computing ability for scientific computing. We will try to address this problem and try to give some mechanisms and software tools in this research.
Each computer system has a certain combination of resources available to it, such as memory, disk, network connectivity, scientific computing function libraies and CPU speed. A resource allocation strategy should allocate processors according to its available resources in order to minimize execution time, subject to the certain availability of scientific functions. In order to fulfill this mission, the task allocator must be able to migrate a tasks from one computer to another during execution.
The task allocator draws upon two sources of information. There should be a database detailing, among other things, the availability of the computer, the informations of scientific function library, the network topology, the habits of the computer's interactive users, CPU performance, local disk performance, and memory size. The task allocator should also monitor performance of the workstations during the execution of the parallel program. This monitor could obtain from the kernel a task's page fault rate, outgoing message rate, message latency, idle time, and memory usage.
Another aspect of the resource management is to explore and add the available computer server or delete the unavailable servers to ths system automaticlly. What we are aiming ro form is a parallel runtime support for global computing based on INTERNET. The computing resource available will be dynamicly changed every day. In order to meet this problem, we should provide the resource management module with this dynamicly change ability.