moskowit@paul.rutgers.edu (Len Moskowitz) (07/07/90)
We are in the process of selecting a processor to be used to implement a small (64 processors or less) parallel architecture. Our uniprocessor baseline is a 68030 running at 20 megaHertz and we would like to better its performance by 20 to 100 times. We will run a parallel production system, perhaps OPS5 or CLIPS. Parallelism will be at the rule or condition element level. This implies that we will not be doing heavy number crunching (mostly integers, no arrays, no vectors). Logical and numerical comparisons will be ubiquitous; hash tables and interprocessor communications too. We have a space problem so compactness is a plus. So far we've identified the transputers, the 88000, and the i860. We are interested in comments and recommendations about the suitability of these and other processors for our task. Len Moskowitz moskowitz@bendix.com moskowit@paul.rutgers.edu
rfg@uunet.UU.NET (Ron Guilmette) (07/10/90)
In article <9589@hubcap.clemson.edu> moskowit@paul.rutgers.edu (Len Moskowitz) writes: >We are in the process of selecting a processor to be used to implement >a small (64 processors or less) parallel architecture. Our >uniprocessor baseline is a 68030 running at 20 megaHertz and we would >like to better its performance by 20 to 100 times. > >We will run a parallel production system, perhaps OPS5 or CLIPS. >Parallelism will be at the rule or condition element level. This >implies that we will not be doing heavy number crunching (mostly >integers, no arrays, no vectors). Logical and numerical comparisons >will be ubiquitous; hash tables and interprocessor communications >too. > >We have a space problem so compactness is a plus. Well, the CPU chips are pretty small... I think that you can get 4 MB SIMMS (and they are pretty small). If you're lucky (and clever) you could squeeze your whole multiprocessor into an average sized broom closet. :-) (As a matter of fact, you can do that now with at least one Intel produced multiprocessor that I have seen.) But seriously folks... is that this what you meant by a "space problem", or were you talking about something else (like code size for instance)? >So far we've identified the transputers, the 88000, and the i860. Be advised that if you want to build a tightly coupled machine, the currently available i860's don't have proper hardware support for maintaining cache coherency. This may or may not be important to you. If all you intend to do is to run numerous relatively independent "processes" (which do not share variables) then cache coherency is not an issue. If you want to run multi- threaded tasks which share variables, cache coherency will be important. There have been rumors that Intel will have an i860 follow on product (someday) that will have proper hardware support for cache coherency, but I would not bet the farm on it until I saw working silicon. -- // Ron Guilmette (rfg@ncd.com) // C++ Entomologist // Motto: If it sticks, force it. If it breaks, it needed replacing anyway.