Multiprocessor Systems<A NAME=1589> </A>



next up previous contents index
Next: Message-Passing Architectures Up: How PVM Works Previous: In the Task

Multiprocessor Systems 

 

Developed initially as a parallel programming environment for Unix workstations, PVM has gained wide acceptance and become a de facto standard for message-passing programming. Users want the same programming environment on multiprocessor computers so they can move their applications onto these systems. A common interface would also allow users to write vendor-independent programs for parallel computers and to do part or most of the development work on workstations, freeing up the multiprocessor supercomputers for production runs.

With PVM, multiprocessor systems can be included in the same configuration with workstations. For example, a PVM task running on a graphics workstation can display the results of computations carried out on a massively parallel processing supercomputer. Shared-memory computers with a small number of processors can be linked to deliver supercomputer performance.

The virtual machine hides the configuration details from the programmer. The physical processors can be a network of workstations, or they can be the nodes of a multicomputer. The programmer doesn't have to know how the tasks are created or where they are running; it is the responsibility of PVM to schedule user's tasks onto individual processors. The user can, however, tune the program for a specific configuration to achieve maximum performance, at the expense of its portability.

Multiprocessor systems can be divided into two main categories: message passing and shared memory. In the first category, PVM is now supported on Intel's iPSC/860   and Paragon , as well as Thinking Machine's CM-5 . Porting PVM to these platforms is straightforward, because the message-passing functions in PVM map quite naturally onto the native system calls. The difficult part is the loading and management of tasks. In the second category, message passing can be done by placing the message buffers in shared memory. Access to these buffers must be synchronized with mutual exclusion locks. PVM 3.3 shared memory ports include SGI multiprocessor machines running IRIX 5.x and Sun Microsystems, Inc.,   multiprocessor machines running Solaris 2.3   (This port also runs on the Cray Research, Inc., CS6400 ). In addition, CRAY and DEC have created PVM ports for their T3D and DEC 2100 shared memory multiprocessors, respectively.





next up previous contents index
Next: Message-Passing Architectures Up: How PVM Works Previous: In the Task