Newsgroups: comp.parallel
From: Mark Rosenbaum <mjr@netcom.com>
Subject: Re: SMP vs. MPP
Organization: NETCOM On-line Communication Services (408 261-4700 guest)
Date: Mon, 23 Jan 1995 12:58:08 GMT
Message-ID: <mjrD2qGCo.42H@netcom.com>

In article <GOTTLIEB.95Jan19102002@allan.ultra.nyu.edu>,
Allan Gottlieb <gottlieb@allan.ultra.nyu.edu> wrote:
>In article <mjrD2GEr3.7Lt@netcom.com> mjr@netcom.com (Mark Rosenbaum) writes:
>
>   >>Finally the $64K question. Does anyone have an algorithm that 
>   >>
>   >a) actually uses writable shared memory for anything other than locks.
>   >
>   I may not have explaned this clearly. In shared memory the systems are
>   defined along two axes read/write concurrent/exclusive. Most algorithms
>   are Concurrent Read Exclusive Write (CREW). I was asking for CRCW.
>
>I am not sure if by locks you mean specifically binary semaphores or
>whether you intended to include all coordination algorithms.  

At the time I was thinking bin sema but notw that you mention it
probabily all control.

>If you meant the former (only bin sema), then I should mention that our
>Ultracomputer group at NYU has devised a number of fetch-and-add based
>coordination algorithms that use CRCW semantics for more than locks.
>For example, queues and readers/writers.
>--

The point is that if you distribute all  information (data & control) &
use some type of FIFO on the messaging to queue request it would
seem to have the same functionality as cache coherent shared memory.

Any comments would be appreciated.

mjr


