Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
8 Achieving a Pioneering Outlook with Supercomputing Lawrence G. Usler Apple Computer; Inc. Apple purchased a Cray-XMP/48 computer a little more than 3 years ago. It has taken a while for us to develop a range of applications. Currently about 20 different projects are using the Cray, and they involve about 50 engineers. Most of the applications that we have are proprietary and cannot be talked about, but fortunately there are some recent applications that I can discuss for this symposium. EXTENDING THE RANGE OF APPLICATIONS lathe main reason we bought a Cray was to make applications possible that were previously impossible because of the time they took to run. We had circuit simulations, for example, that would have run for 2 months, and it was easier to actually build the circuit and try it out than to wait the 2 months to run the simulation on, say, a VAX. The Cray has helped us a lot, because now we can run those simulations in 1 day. We had applications that would run overnight, and you had to really plan them well, start them running, and then come back the next day to see the results. If there was a mistake, you had to make a little change and run the application again. Those we can now do in a few minutes, and we can try many variations quickly, as the other speakers have mentioned. More importantly, there were things that previously had to be done in the batch mode that we can now do interactively, a requirement that Beverly Eccles of Abbott Laboratories talked about very well. But the main 90
ACHIEVING A PIONEERING OUTLOOK WIIU SUPERCOMPl]TING 91 reason we bought the Cray is that we ourselves are a computer company, and our job is to create the computers of the future. The group that I run, the Advanced Technology Group, is not de- signing products for Apple. We are doing research on and prototyping of technologies that will apply to future products. So we are looking 3, 5, or 10 years ahead for what might be applicable in future Apple products. For us, the supercomputer is a way to experience the kinds of speed that will be on the desktop in the $1,000 to $10,000 range in several years. 1b make that possible, we have created a somewhat unusual setup. The network that we have at Apple includes the Cray-XMP. We recently upgraded the disk storage on that to about 30 gigabytes. In addition we have an EN-641 that connects us to Ethernet, and we have VAXs and many Sun and Apollo workstations on Ethernet as gateways into the AppleTalk network, which allows us to connect up to the many thousands of Macintoshes that are all around the Apple campus, including more than 1000 in engineering. On the Macintoshes we have software, for example, NCSA-TELNET, as well as a product from Pacer software that allows us to do terminal emulation, file transfer, and so on, by using convenient menus. We are able to produce not only text but also graphic displays on the Macintosh to access the power of the supercomputer. We also have, in addition to the standard 50-megabyte hyperchannel, an 800-megabyte-per-second ultrachannel that gives us very high bandwidth video out, essentially, to a number of high-resolution monitors, so that we can get direct interaction with the Cray. When we do that, we are temporarily tying down the entire machine for one user. If someone is rotating an image in three dimensions, the machine is dedicated as a personal computer to that user for a few seconds while that is going on. Some of the applications I will discuss rely on that capability. We use the Cray both in product development and in research. We use it for circuit design simulations, and that has saved a lot of time in proving designs. Disk head design is an application I will discuss; industrial design is another. The disk head design project is an interesting one. The goal is to make the recording head fly at a constant height over the disk surface, on the order of 10 microinches. The shape of the head and the shape of the medium and the aerodynamics all interact. If the system is not set up well, then the head will crash, or the head will be too high above the disk to get a clean signal. We wanted to know what the effects of various parameters were. An interesting problem we ran into was that if the head itself, the air bearing, has a resonant frequency that corresponds to the frequency of the rotation, oscillations result. To understand that better, we worked with Jim
92 LAWRENCE G. TESLER White from the University of Santa Clara. The disk head can have little bevels and other shapes that affect the aerodynamics. Our engineers came up with a set of equations based on the geometry of the head. One of the problems is that the medium itself is not completely flat. It can be a thousandth of an inch off flatness, and when the head is flying a few hundred thousandths of an inch over the disk, that variation can cause a problem because essentially, the head is like a cruise missile trying to go over peaks. In a simulation, the head can be shown as flying between-150 and 350 or so nanometers over the surface. By varying the shape of the head, the engineers can run the simulation over and over. The simulation takes only 20 minutes to run, and the engineers can keep playing with it until they achieve satisfactory results. Another concern is that there may be a problem caused by a slight bump in the medium. A little of the oxide may have a small bump in it, and a result may be that the head can really bounce. And of course if it bounces too much, it will crash. Another area to explore is what happens if there is a jolt to the head, which can happen because someone moves the drive while it is running. After a seek, when the head comes to a stop, there is a similar jolt. We need to know how long it will take the oscillation of the head to settle down so that we can actually start to do a read or write. Why is Apple studying all these things when in fact we do not man- ufacture disk mechanisms? The reason is that we work with vendors of heads and media, and vendors of drives, and they come to us with claims of why the next generation of disks is going to be so much better than the last. We need to be able to evaluate their claims, because if we simply go along with them and something does not work well, then we might have to shut down our production, and that is a serious consequence. Product Design We have also used the Cray for product design, or packaging. We have used three different applications: (1) thermal analysis, (2) structural analysis, and (3) mold flow similar to that discussed by Cliff Perry. For thermal analysis we have used a package called ANSYS, which is a finite element program displaying the output graphically on a Macintosh II. For example, we have modeled a personal computer board, with major heat sources displayed as small blue areas. A simulation is run until it settles into a steady state, which occurs after the computer has been on for a while. Then the task is to see what temperatures the various components have reached. ~,
ACHIEVING A PIONEERING OUTLOOK WITH SUPERCOMPUTING 93 It is possible to tune the heat flow in the system by adjusting the cooling air how and the layout so that it is all in the range of about 143 to 152°F, but some hot spots may remain. By playing with the parameters, engineers can try to get the temperatures within an acceptable range for the components. Another application, called NEKTONics, is a finite element program that is used for structural analysis related to the cooling problem. A small object represents the edge of a cooling vent. Just above and below that object is a vent; a piece of plastic separates the vents. As air is drawn in through the package, the flow is depicted. The question is, What will the temperature be after a certain period of time, given certain assumptions about air flow and the initial conditions? This program can show potential problems. For example, if air flows past a particular point and loses velocity, it also loses its ability to cool. It is possible to play with the shape of a particular edge and solve the problem of reduced air flow. By manipulating with a computer-aided design (CAD) program the shape of a vent edge, a different flow pattern was obtained, and the result was that the velocity loss was reduced. A third application is used for a mold flow problem. What is interesting in this example is that the product we used this application for was the large-size, extended keyboard for the Macintosh II. A keyboard for a Macintosh has various places on the surface that, if looked at in just the right way, are small dark areas. These are weld lines where the plastic has come through the mold and welded together, and they are not very good to look at. The problem was to try to improve the keyboard's appearance so that people would stop telephoning Customer Support to ask why they couldn't clean their keyboards. The approach to this problem was to break the keyboard's surface down into very small polygons and then to run a simulation that showed the filling of the plastic in the mold. The point at which the plastic in two paths merges together becomes a weld line. The idea is to try to control conditions so that the temperature of the two is about the same and the weld occurs in a place where it will not be noticed by the user. This is the case in the newly designed keyboard, not in our original design. In a two-dimensional display of mold flow, different colors represent different time periods so that it is possible to see the history of the flow. Now there is an interactive program that shows the process in real time. The engineer can use a mouse to select a specific part of the picture and then can view a blown-up zoomed-in view of just that part. This gives the engineer the ability to focus on parts of the process. Our application does not have the aesthetics of the visualization that Cliff Perry described or the ability to show multiple parameters at once. Instead, we traded that off to be able to get interactive capability for the engineer.
94 LAWRENCE G. TESLER We also can do pressure and temperature plots, and so on. What is important is that in the end, the engineer gives to the plastic maker a drawing that shows the key things that have to be done. Basically the approach to solving the problem of getting the plastic to flow at the rate that we wanted was to indicate places where the inside of the mold was narrower, which slowed down the plastic flow so that we could catch up in other places. This approach gave the result we wanted; the weld lines were exactly where we wanted them. The illustration was done with a program called PIXELPAINT on the Macintosh II, by starting with the CAD diagram and simply taking the data that came out of the simulation. The benefits of using the supercomputer for product design are in- creased savings of time and money hundreds of thousands of dollars- made possible by fewer tooling runs plus the much greater advantage of getting products to market faster. We can get products to market months faster because we know that the likelihood that the first mold is going to work is much higher. We also do not have to wait months for another mold if there is something wrong with the first one, and we can improve the various parameters of the design and get better quality. Research Now we also use our supercomputer in research. At Apple, we have been doing neural network simulations to better understand how to use different neural net models for learning. In addition, we have simulated a cochlear model that is used in a speech recognition project. The idea is that, to be usable, any speech recognition system has to be able to work in a noisy room. One approach to achieving that is to try to actually model the human ear, which has a comb of hairs that is able to sort out different frequencies and to measure, essentially, the energy at each different frequency. Some work had been done at Schlumberger Research by Dick Lyon, who recently came to Apple. What we decided to do was to take the same type of model that he had implemented, implement it on the Cray, and then animate the results. The result is a plot, called a correlogram, that shows low frequencies at the top and high frequencies at the bottom. Viewed from left to right, it shows various correlations of different timing sets. Essentially it enables the engineer to "see" what the ear "sees" when it hears a sound. An utterance can be visualized as pillar-shaped forms that represent the main frequency and as other forms that represent other, weaker frequencies. If there is noise in the room, the background becomes fuzzy, but it is still possible to see a pattern of frequencies standing out. This is the beginning of a very
ACHIEVING A PIONEERING OUTLOOK WIHI SUPERCOMPUTING 95 long research project to try to emulate the power of the human ear to sort out noise. For neural net simulation we have been able to get very high rates on the Cray. Using even only one processor on the XMP, we have been able to do about 10 million connection updates per second and also to animate the results to get a feel for how a neural net learns. So the benefits for research are, again, more rapid prototyping. We can try many alternatives. People are willing to try things if they can get results in a few minutes, or even interactively, but the main thing is that we are now much bolder. We will try things that we would not have tried before. One of the chips that we are designing currently is one we probably would not have attempted to design previously because people thought it would not work. When we simulated it, we found that it would work, and we went ahead and built it and in fact it did work. So I would say that the main impact of the supercomputer is that it makes us more comfortable with taking bigger risks. ADDING SUPERCOMPUTING CAPABILITY Visualization, as everyone participating in this symposium has ex- plained, is an absolutely key capability. Having a fast network is very, very important, and we continue to upgrade the speed of our network so that people can get higher bandwidth between the user interface and the supercomputer. One big problem is simply operating the supercomputer center. It accounts for a major portion of our budget, and we are always under pressure to add new power to it. It is competing always with other needs such as upgrading the network and adding minisupercomputers and workstations. The operations end is something that anyone thinking of buying a supercomputer really must consider. Two years ago we brought in a person from our Management Infor- mation Systems Department to manage our supercomputer center, and we have hired several people who are expert in using the CraY and other supercomputer engineering systems to work there. The last hurdle, as other people have mentioned, has been to get the users to use the supercomputer. The way we do it is that we have a small group of people we call Cray evangelists. They do not appear on television; they walk around. They go to engineers and try to find out what those engineers do, and then they try to match them up with applications on the Cray or help them write their own applications on the Cray. All of the applications I have discussed in this symposium have come from that effort, which is very similar to what was described as the effort that goes on at Kodak also.
96 DISCUSSION DISCUSSION Edward Abrahams: Have you who have tackled this problem of con- vincing users to use the supercomputer found some techniques that were not so productive? Presumably you have mentioned some of the ones that are productive. What techniques did not work, so we can avoid them? Lawrence Tesler: Trying to convince somebody who is very negative is probably the one thing that isn't worth doing. In other words it's important to find people who immediately see the benefits of supercomputing and to get them to start using it. Then their colleagues will realize that they can use supercomputing also. Beverly Eccles: Yes. Seek the champions for the cause. Clifford Perry: I don't have too many keys to failure, but one key to success is involving the users from the very beginning in participative planning. We actually sent out letters to literally hundreds of the heavy users of our traditional high-end mainframe computing facility, asking them to participate in an idealized design of a supercomputing facility and to think about what the attributes of that particular center should be. Would it offer one-on-one collaborative assistance? Would it offer transparency vis-a-vis using that computer or the high-end mainframe? How would it be administered? How would it be charged out? What help would be rendered to the users? When only top-down decisions are made, people don't use the comput- ers. The decision-making process has to be top-down, bottom-up, middle- out. We focused on the bottom-up and middle-out, and then when Larry Smarr came and mapped what he had to offer against the idealized design that was documented and was formulated by the participation of those whose lives would be affected by the advent of the supercomputing facility, we found that we had an immediate, captured market. Generally, it takes about 2 years to justify the use of a supercomputer onsite, and it takes upwards of $250,000 to $300,000, as has been published by the Minnesota Supercomputing Consortium. It has to be done in a participative manner, in my opinion, or it won't work. Mel Schmidt: Could all the panelists briefly describe how they deter- mine allocations within their organizations? Are there any mechanisms for billing the users? Beverly Eccles: Within Abbott, computer resources are basically free. The resources are supplied and the expense goes into the budget, but individuals are not concerned about how much disk space or how much of the central processing unit they are using. Those resources are simply available for us, and the scientists feel very comfortable in that environment.
DISCUSSION 97 Clifford Perry: At Kodak we have an arrangement with NCSA that every user who logs on-and we have an administrative procedure to do that-is billed directly in their division. We have allocated, if you will, $100,000 chunks to sets of people. Lawrence Tesler: At Apple as at Abbott, all the shared computer resources that are used by more than one department are essentially free. On our financial reports from the Apple Product Division, we break out the entire budget for this computer operation. It is weighed as a whole as a percent of the entire R&D budget.