Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!sdd.hp.com!ucsd!ucbvax!sprite.Berkeley.EDU!elm From: elm@sprite.Berkeley.EDU (ethan miller) Newsgroups: comp.arch Subject: Re: LINPACK 1000x1000 MFLOPS per $$$ Message-ID: <37683@ucbvax.BERKELEY.EDU> Date: 20 Jul 90 22:53:15 GMT References: <2349@crdos1.crd.ge.COM> Sender: usenet@ucbvax.BERKELEY.EDU Reply-To: elm@sprite.Berkeley.EDU (ethan miller) Organization: U.C. Berkeley Sprite Project Lines: 98 In article , mccalpin@perelandra.cms.udel.edu (John D. McCalpin) writes: %In article <> mccalpin@pereland.cms.udel.edu I wrote about MFLOPS/$: %The configuration that I quoted has a rather small memory by current %supercomputer standards, but 2 MW (64-bit) is hardly "tiny". Problem #1 shows up right now. You compare price/performance for the $13000 machine, and then turn around and say that you'd actually need to pay two or three times that for a machine that will do what you want, or even what you claim. %Estimated cost for a full 128 MB = 16 MW is about %$20,000 in addition to the base price of $8700 for the machine. So let's assume you only get 8MW of memory, at a total price of about $20000 for the machine. You've cut your advantage in half, and all you've done is buy more memory. Now start buying enough disk space for that simulation data.... %Since Cray is still selling lots of Y/MP's with 32 MW memories, it is %hardly fair to criticize a single-user workstation on that account. Cray's I/O system is able to handle much more "paging," where programmers shuttle data in and out to fit in a tiny memory space. Can the PowerStation accommodate this? It's not just a question of I/O bus bandwidth; the disks must be able to keep up as well. %As far as disk storage goes, I have 1.5 GB of disk space on my %graphics workstation, and will soon have a 2.3 GB tape drive. So %manipulating 500 MB datasets (see below) is entirely practical. % %The whole setup: %[... setup deleted; see referenced article] %is under $50,000 at University prices. So now we're up to $50000 for the configuration that you're racing against a Cray. Suddenly, the killer micro isn't as killer. Of course, if all you want is lots of MIPS and MFLOPS, and you don't need much memory, you're still OK. However, the original cost/performance ratio has just dropped by 4 or 5 times because you've added enough components to make a real system. % (1) The cost must be within the available budget. % This includes the cost of porting the code as well. Is it any harder to port code to the Cray than to other machines? How about other supercomputers, such as Convex? There will certainly be porting costs, but I don't think they'll be much worse for a supercomputer than for any other computer. Please correct me if I'm wrong on this, though. % (2) The wall-clock turnaround must be within the limits % of the research project. If you suffer a 25 to 1 slowdown of CPU time, that will change turnaround times from overnight to one month. That's a big difference. % (3) Point (2) usually requires sufficient memory to make % the problem core-containable. Not necessarily, especially if you're running on a computer with lots of I/O bandwidth (assuming you have the devices to feed it). There are also quite a few simulations that aren't core-containable on any Y-MP. What then? % (4) Sufficient mass storage space and access speed must be % available to save intermediate and permanent results % without slowing down the calculation past the constraints % of point (2). It is this element that can contribute lots of cost to a computer system. % I recently submitted a proposal to the NSF to do some cpu-intensive %studies of the equations governing a theoretical two-dimensional %ocean. The calculations are estimated to require 200 hours of Cray %Y/MP time. I don't consider this a trivial expenditure.... %With an IBM 320, I would probably be able to finish all of the %calculations before the proposal even completes the review process! Really? That's 5000 hours on a PowerStation (using the 25/1 ratio from the table). That's about 200 days, assuming you use every single CPU cycle on the machine. Since you'll be doing some I/O, though, I'd be surprised to see better than 50-75% utilization, which brings total running time to close to a year. Granted, it's cheaper than Cray time, but is it practical to wait a year for a single simulation to finish? There are simulations that can run on a workstation instead of a supercomputer. These tend to be smaller simulations, though, for memory, disk/tape storage, and CPU speed reasons. You can probably increase one, perhaps two dimensions of these axes and stay with a workstation. Once you increase all three, you have a supercomputer, or pretty close to it. ethan ================================= ethan miller--cs grad student elm@sprite.berkeley.edu #include {...}!ucbvax!sprite!elm Witty signature line condemned due to major quake damage.