terry@uts.amdahl.com (Lewis T. Flynn) (05/18/91)
There was a discussion in this group recently concerning Amdahl's Rule and whether it should be updated, if it still held, etc. Some portions of what was being quoted didn't ring exactly true so I did some homework. First, I checked Hennessey and Patterson to see what was said on the subject there and it mostly agreed with what was on the net. So rather than do more second hand research, I called Gene to ask him about it. As it happened, he was out of the country, but, being the gentleman he is, he returned my call when he got back. As to the 1 megabit of I/O capacity per MIPS rule: he and his people did a lot of research on what workloads were being run on the then current machines (709/7090) and found that, in general, these workloads used 1 bit of I/O per instruction. Thus, in order to run these workloads, you must be able to sustain this rate. Peak I/O capability must be much higher to handle variations in demand. The actual data supporting this was from several sources and included a 2 month study at Lawrence Livermore Laboratories. The numbers from Livermore were 1.01 bits per instruction during the day and 1.1 to 1.2 bits per instruction for night time processing (my notes are a little fuzzy, so I may have reversed the day versus night). He didn't say so, but from the time frame and from the extensive study done at Livermore, I would speculate that this work was done as part of the research for the Stretch project. As to the 1 megabyte per MIPS rule: he stated that this was influenced by the cost of memory for the 709 series. He later found that on the 7090 that 2 megabytes/MIPS was more reasonable and this value held when he was doing 360/370 design at IBM and Amdahl. He said it really showed when the 370s hit the 16meg limit (24 bit addressing) and faster processors got no better throughput. I guess the most useful thing to conclude from this is that the first rule of processor design is "Know Thy Workload". The second most useful thing is that there's no substitute for lots of research and measurement. Terry Disclaimer: Dr. Gene Amdahl no longer has any affiliation with Amdahl Corporation and my discussion with him was completely unofficial. I had met Dr. Amdahl in the course of my duties at a previous company.
wigen@netcom.COM (Patrick L. Wigen) (05/19/91)
In article <48qO01yL75GS00@amdahl.uts.amdahl.com> terry@uts.amdahl.com (Lewis T. Flynn) writes: >There was a discussion in this group recently concerning Amdahl's Rule ... >I guess the most useful thing to conclude from this is that the first >rule of processor design is "Know Thy Workload". The second most useful >thing is that there's no substitute for lots of research and >measurement. > >Terry I do alot of work with many different systems mostly workstations lately and whiel most people recommend roughly following Amdahl's rule (and are probably not aware of it) I have noticed that doubling that figure is usually most efficient. For example a Decstation 3100 that runs at ~12-14 MIPS running decwindows (X) runs much better with 24 MB of memory than the usual recommendation of 16 MB. This is just a casual observation hoewever and is not based on emperical data. Patrick L. Wigen wigen @netcom.com no .sig yet.......
jerry@talos.npri.com (Jerry Gitomer) (05/21/91)
wigen@netcom.COM (Patrick L. Wigen) writes: :In article <48qO01yL75GS00@amdahl.uts.amdahl.com: terry@uts.amdahl.com (Lewis T. Flynn) writes: ::There was a discussion in this group recently concerning Amdahl's Rule :... ::I guess the most useful thing to conclude from this is that the first ::rule of processor design is "Know Thy Workload". The second most useful ::thing is that there's no substitute for lots of research and ::measurement. :: ::Terry :I do alot of work with many different systems mostly workstations lately :and whiel most people recommend roughly following Amdahl's rule (and are :probably not aware of it) I have noticed that doubling that figure :is usually most efficient. For example a Decstation 3100 that runs at :~12-14 MIPS running decwindows (X) runs much better with 24 MB of memory than :the usual recommendation of 16 MB. This is just a casual observation hoewever :and is not based on emperical data. Since most "natural" workstation applications tend to be processor intensive while general purpose computing applications (the kind that I understand Amdahl based his rule on) tend to be I/O intensive this makes sense. This only goes to reinforce the need to "Know Thy Workload". -- Jerry Gitomer at National Political Resources Inc, Alexandria, VA USA I am apolitical, have no resources, and speak only for myself. Ma Bell (703)683-9090 (UUCP: ...uunet!uupsi!npri6!jerry )
ddr@cs.ed.ac.uk (Doug Rogers) (05/23/91)
Amdahls law on memory must be thought of as pragmatic. There is no simple causal relationship that would suggest this ratio of between 1 and 2 Megabytes of data per MIP. Two effects are at present anyway pulling in opposite directions. The use of RISC machines demands more memory and the use of shared libraries reduces memory needs. I think the DEC machine does not use shared libraries as is the case for the HP offerings at present. Both these machines therefore require an excessive amount of memory compared to the SUN machines. The size of memory is anyway dominated by the memory access time versus swap time. Speed up all the parts of the machine simultaneously and the machine will run proportionaly faster. Here I think the introduction of file servers and/or the use of discs which are not significantly faster than 5 years ago has forced the holding of all the cde in memory. ie page fualting is now very expensive. The argument of cost also does not work as the relative cost of the system components has been changed dramatically. I wonder if the reason the ratio is being kept is because it is being designed for? -- Douglas Rogers JANET: ddr@uk.ac.ed.lfcs Department of Computer Science UUCP: ..!mcvax!ukc!lfcs!ddr University of Edinburgh ARPA: ddr%lfcs.ed.ac.uk@nsfnet-relay.ac.uk Edinburgh EH9 3JZ, UK. Tel: 031-650 5172 (direct line)
lindsay@gandalf.cs.cmu.edu (Donald Lindsay) (05/24/91)
In article <11299@skye.cs.ed.ac.uk> ddr@cs.ed.ac.uk (Doug Rogers) writes: >The use of RISC machines demands more memory and the use of shared >libraries reduces memory needs. Surely the ratio of code-to-data shifts towards data as memory size increases? I would have thought that this swamped the effects you mention. -- Don D.C.Lindsay Carnegie Mellon Robotics Institute
ddr@cs.ed.ac.uk (Doug Rogers) (05/24/91)
In article <13179@pt.cs.cmu.edu>, lindsay@gandalf.cs.cmu.edu (Donald Lindsay) writes: > In article <11299@skye.cs.ed.ac.uk> ddr@cs.ed.ac.uk (Doug Rogers) writes: > >The use of RISC machines demands more memory and the use of shared > >libraries reduces memory needs. > > Surely the ratio of code-to-data shifts towards data as memory size > increases? I would have thought that this swamped the effects you > mention. > From our experience of university computer usage, I would disagree. This was shown clearly when the SUN OS 4.0 came in, and a teaching cluster was transformed from unusable to usable. I don't have any figures though on this. -- Douglas Rogers JANET: ddr@uk.ac.ed.lfcs Department of Computer Science UUCP: ..!mcvax!ukc!lfcs!ddr University of Edinburgh ARPA: ddr%lfcs.ed.ac.uk@nsfnet-relay.ac.uk Edinburgh EH9 3JZ, UK. Tel: 031-650 5172 (direct line)
perry@hpfcdc.HP.COM (Perry Scott) (05/25/91)
>Re: Amdahl's Law > >The use of RISC machines demands more memory and the use of shared >libraries reduces memory needs. I think the DEC machine does not use >shared libraries as is the case for the HP offerings at present. > >Douglas Rogers Just a minor nit - the 8.X releases which are just hitting the streets have shared libraries. This is true for the current 9000/700 releases as well. Perry Scott HP Ft Collins