Path: utzoo!attcan!uunet!tut.cis.ohio-state.edu!zaphod.mps.ohio-state.edu!usc!elroy.jpl.nasa.gov!jarthur!nntp-server.caltech.edu!chamber!jerbil From: jerbil@chamber.caltech.edu (Joseph R. Beckenbach) Newsgroups: comp.sys.amiga Subject: Re: Mach on Amiga Message-ID: Date: 28 Jun 90 16:33:38 GMT References: <1990Jun27.003638.27525@zorch.SF-Bay.ORG> Sender: news@laguna.ccsf.caltech.edu Organization: California Institute of Technology Lines: 40 xanthian@zorch.SF-Bay.ORG (Kent Paul Dolan) writes: >Yes! I have no idea of the Mach internals; if it is a "supervisor and >subordinates" paradigm, then lots of cheap coprocessors on a plug in >board might be doable. If it is a "many peers" paradigm, then boardfulls >of 68040's could run less tightly coupled. Either way, talk about your >screaming graphics! Where's that tower with all the slots!?! There are three ways to organize parallel processing hardware: a processor shares memory with other processors (global memory), a processor does not share memory with other processors (distributed memory), and a hybrid of the two. (For a small number of processors (<16) the global memory scheme is okay; about 128 processors, and distributed-memory starts looking really good. The distributed-memory scheme scales quite easily through thousands.) Mach assumes a shared-memory machine, and can tolerate being run on a hybrid machine. Local-memory machines, such as anything based on the Connection Machine, or on Caltech's Cosmic Cube (Intel ipsc's, Symult 2010), or on transputers [:-b] would be able to run Mach; however, they would each need to run on top of a global-memory simulator to supply the conditions for which Mach is suited. This is a major performance and parallelism loss for the distributed-memory machine. Mach on the Amiga would work well for however many gangs of processors you wish to cluster onto a given memory pool. In fact, a collection of four processors and shared memory on a board, which can use the Amiga buses as a network, would work well. I personally am much more interested in the distributed-memory cases, since I want to eventually use an Amiga as an interface into a box with 1024 (or so :-) processors. Lots of memory, lots of computing cycles. Lots of power :-). Ah well, back to the coal mines. Joe Beckenbach Caltech CS department jerbil@csvax.caltech.edu joe@vlsi.caltech.edu