schow@bnr-public.uucp (Stanley Chow) (02/16/89)
In article <3682@geaclib.UUCP> daveb@geaclib.UUCP (David Collier-Brown) writes: > [David would like cheap D-Cache kits so he can build large caches]. > I want a machine with performance somewhat better than my >Stunned despite my stuffing it to bursting with slow main-memory >chips. I want an architecture which will deal with large working >sets and ill-behaved (not very local) reference patterns without >having to sweep the slowdown under the rug by running large numbers >of users/processes in a timesharing system. > I'd like the opportunity to do this at finite cost in the >near future by allowing me to bite the bullet and put in lots of >cache in front of a rather huge but not-too-fast main memory. I >really do want lots of memory, and am betting that the difference in >memory speed/cost versus processor speed/cost makes investing in >cache worthwhile. > > The assumption I'm making, you see, is that it's worthwhile >to allocating particular cache chips to particular blocks of memory >chips. Probably along paging boundaries. This in turn means that >I'm betting that cache invalidation/reloading on process change is a >major bottleneck when one has large amounts of cache. This in turn >means that I'm still betting on simple multiprocessing for this kind >of application domain. If you application has large working sets (I assume you mean many times larger than a reasonable D-Cache - say 256K bytes), with ill-behaved reference patterns, why do you think D-Cache will give you performance with slow main memory? For a large system, memory and I/O will often limit performance purely on economics: the CPU can be very fast, but you can't afford the fast memory or fast disk (or your customers won't buy). If you are willing to go to the trouble of dedicating cache sections for particular process/memory, why not turn the cache into simple SRAM and statically allocate the SRAM? If you mean partition the cache on a per process basis so that you can save on proces change time, I would think you will run out of cache partitions pretty quick or have pretty small partitions. Please, no dicussions on RISC/CISC suitability. This is strictly a question on the Data side of things. Stanley Chow ..!utgpu!bnr-vpa!bnr-fos!schow%bnr-public (613) 763-2831 Disclaimer: Just because I am working on architecture and processors and stuff for a living, please don't think that my employer knows anything about my opinions. (Or that they would care anyway).