Date: November 10, 1994

Speaker: Thomas Stricker

Twin Peaks: A better use of a processor/co-processor pair in distributed memory parallel computers

Abstract:
Both schools of parallel system architects, shared memory and message passing, agree that the key to good communication performance in parallel programs is flexible, fast block transfers. There is a significant processing overhead involved with sending and receiving such blocks of data into/from a high speed network. The designers of parallel machines therefore have added more computation power in the form of a co-processor to the nodes hoping to help with the protocol processing [Nectar, IBM SP/2, Paragon, Parsytec, SPLASH]. In this model, the main processor performs the tasks of the higher layers of protocol (e.g. coallescing data, computing the communication schedule) while the co-processor performs the lower layers of the protocols (e.g. calculating the routes, flow control, and injecting / extracting messages from the network).

From my experiences with the Nectar and iWarp message passing systems, I noticed that much of the communication overhead comes from dealing with the non-determinism of the network. This makes handling communication an inherently multithreaded task, where processing is switched rapidly between the sender and the receiver threads.

We are proposing that communication steps in parallel programs are better performed by a single-threaded sender and a single-threaded receiver, running on the processor and co-processor respectively, rather than by two multithreaded programs doing work for sending and receiving simulanously, splitting work between the high and low level protocols.

I believe that multithreading will never be a good idea for high performance off-the-shelf microprocessors, and I will outline two architectural reasons why "twin peaks" should work better than the traditional processor / co-processor setup.

I will present preliminary supporting experiments from iWarp. We plan to implement "twin peaks" on the Paragon using the Sandia SUNMOS environment. This will be joint work with Susan Hinrichs and Michael Hemy.

SDI / LCS Seminar Questions?
Karen Lindenfelser, 86716, or visit www.pdl.cmu.edu/SDI/