yba@arrow.bellcore.com (Mark Levine) (09/07/88)
I am interested in obtaining a survey of modern and expected-soon bus speeds. I would be interested in hearing from anyone who has a good survey, knows where to find a list of bus architectures and transfer rates, etc. My purpose will be to compare moving bits across a computer bus with other paradigms for moving computer data about (like a Local Area Network). Bus limit(s) imposed by DMA hardware, maximum burst rates, or contention is useful information. (Example: folklore has it the VME bus could go at 160 MB/sec if ALL the controllers were designed to cooperate, but practically VME always goes at a small fraction of that). How bold a statement is "the 60 MB/sec I/O throughput rate of a DEC 6200 is about as fast as we can expect to see for a few years in the workstation/mini/super-mini class" (whatever that class is, mostly under $1M)? How many things are out there reliably pumping over 150 Mb/s? Are manufacturer's quoted speeds inflated by the width of the bus, number of processors, or assumptions about contention? Quotable sources appreciated, but not required. Is this an appropriate place to ask? Any better pointers?
jim@belltec.UUCP (Mr. Jim's Own Logon) (09/08/88)
In article <849@sword.bellcore.com>, yba@arrow.bellcore.com (Mark Levine) writes: > > > My purpose will be to compare moving bits across a computer bus with > other paradigms for moving computer data about (like a Local Area > Network). Bus limit(s) imposed by DMA hardware, maximum burst rates, > or contention is useful information. (Example: folklore has it the VME > bus could go at 160 MB/sec if ALL the controllers were designed to > cooperate, but practically VME always goes at a small fraction of > that). > > How bold a statement is "the 60 MB/sec I/O throughput rate of a DEC > 6200 is about as fast as we can expect to see for a few years in the > workstation/mini/super-mini class" (whatever that class is, mostly > under $1M)? How many things are out there reliably pumping over 150 > Mb/s? Are manufacturer's quoted speeds inflated by the width of the > bus, number of processors, or assumptions about contention? Quotable > sources appreciated, but not required. > Perhaps the simplest way to aproach this is to look at the factors involved, as opposed to what has been done already. The simple factors in bus transfer bandwidth is the bus width and the transfer time. Assume 32 bit bus (although for a multiprocessor memory bus I would prefer 128 bits) and assume that the majority of transfers are from a cpu to memory. At this stage in the hardware evolution, the processors are much faster than the memory. so all transfers are memory bound. For 100 nsec RAM, the fastest complete cycle will take about 200 nsec. This is 5 million transfers a second. If every transfer moves 32 bits this is a throughput of 160 megabits per second. This is about what VME and MB II quote for their burst transfers. It would be easy (relatively speaking) to design a general purpose bus to do double this. THe hardware to achieve this transfer rate would be rather expensive though. This has been a very general answer to a very complex question. Their are a lot more factors involved than those I have mentioned (so nobody jump all over me for simplifying). This is what us hardware architects get paid big bucks to do. Weigh all the factors, costs, desires, and problems and make a compromise that annoys everybody. -Jim Wall