Click for next page ( 120


The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement



Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.

OCR for page 119
Session V Summing Up Robert B.K. Dewar Allen C. Larson R.A. Young ll9

OCR for page 119
SUMMING UP Dewar: I will start with a comment on the qualitative estimates Walter Hamilton presented in his opening remarks of what expan- sion was expected in the field over the next five or ten years. Well, it depends on the number of graduate students, number of postdoctorals, and other unknowns, but nowhere was there a sug- gestion that it might depend on a number of worthwhile problems. I think that significant because I think a sociological change has occurred forcing us to look at our research programs in a different way. The way we shall have to follow is to determine that there are certain problems to be solved, to assess the costs associated with their solution and then estimate whether the solution is worth the associated costs. We have seen very little of that kind of reasoning in the last decade, because when one's expanding rapidly one doesn't need to answer these questions. But the age of rapid expansion in some ways is certainly slowing up, and I think the impact of this on the computer question in particular is that we must be very much concerned with using computer resources effi- ciently. My entire impression from many things said today is that we are not doing this now and we do not have a means for doing it. For instance, we do not have rational systems for spending mon- ey. In most situations, the computer is sort of cream off the top which has to be scrounged from money in a different category from that provided for other expenses. I have made an estimate, which may be disputed and not entirely accepted though I believe it to be approximately correct, that a professional research scientist costs about $40 000 a year. That is the kind of figure non-profit and industrial research organizations assess. Then it is worth spending quite a few dollars of computer time to optimize utiliza- tion of personnel resources, whereas at the moment we have no mech- anism for rational division of funds among the various costs. It really doesn't make much sense to talk about computer costs per structure. What we want to get at is the cost per structure and a rational way of minimizing that. I have a further specific observation. I have a feeling informa- tion interchange is missing. There seems to be a lot of information about possibilities of computer hardware and computer software, and that information is not being transferred effectively within the crystallographic computing community. One of the things I hope would come out of this meeting is perhaps not a specific recommen- dation, say some network to use or some gigantic computer to be built, but that continuing committees should look into the whole question of efficient utilization of resources. I do not current- ly see any mechanism for doing this in an effective way. 120

OCR for page 119
Larson: Mr. Wagner of the AEC has informed us of the possibility that the AEC will acquire several new computers in the near future. Al- though he has suggested that it may be possible to buy time on these facilities, it may be very difficult. From what has been said, com- puting networks like the ARPA Network will come into being, but from what I have heard listening to Division of Military Applications per- sonnel, I seriously doubt that many of the government laboratories, particularly the two with the largest computing arrays, will get into providing external services. If people are interested in trying to use the computing power that is available at the AEC laboratories, it might be worthwhile to contact the staff at these laboratories and see what can be worked out. But I would not hold out a good chance of success. Walter Hamilton in his survey attempted to find out how much it costs to do a structure, or what the kinds of costs in doing struc- tures are. I am going to interject some observations of my own from studies in our laboratory. I have written a code that comes moder- ately close to doing what Bill Busing's AXI0M does. I can set up and do the data reduction, go through the symbolic reduction with- out having given to the code anything more than the numbers neces- sary to control the processes. This comes out at the end of a short run with the structure-factor table printed and ready for publication, along with a stereo drawing of the structure. On one simple structure, in fact one of the two that D.T. Cromer talked about (at the ACA meeting preceding the conference), I believe the computing cost was about five dollars. That has nothing to do with the costs of collecting the data, but gives an idea of what happens with a code like AXI0M at a large computer center with a fairly so- phisticated set of programs. It is quite obvious that we are going to have to go to some kind of network. We are going to have to use large computers for many of the problems, particularly in the area of refinement which Dr. Johnson brought up. I recently asked our local systems people to set up so that I could easily do seven-hundred parameter least- squares on our equipment. They asked, "Are you sure this system is non-singular? Are you sure this system will not blow up?" I per- formed the tests they asked me to, and they then said, "This really is a stable system." Then they said, "We are proposing not to solve the least-squares problem for you in the normal procedure, but rather to solve it without formation of the normal equations." And this, of course, is a fairly reasonable procedure. You do not end up with all of the sums of products. It will probably not speed up the solution but we will not be playing with all of these squares. This procedure is some form of the Householder method; I don't know much about it but it is something our better theoreticians might take a close look at. 121

OCR for page 119
With regard to trying to do structures with mini-computers, I think this is obviously the way for a lot of small laboratories to go, or at least seriously to think about. Because as Sparks said, Vitamin Bj^ w^s solved on a computer that does not have the power of a PDP-8 or a NOVA. We are leaving a lot of our computing power sitting there doing nothing. I have been unhappy about that sort of situation in my laboratory for a long time, but have been too busy trying to make the big computer do all my work to go down and use the little one. Young: What was the $5 structure? Sodium chloride? Larson: It's the azobiscarbamide Cromer talked about. It's a pair of nitrogens sitting across a center of symmetry with a carbamide group attached to the nitrogen at each end of the azo chain. This is a simple structure. Hamilton: You mentioned AXI0M. There may be a number of people here who weren't at the ACA banquet the other night who don't know what AXI0M is. Larson: AXI0M is a procedure where you put the crystal on the diffract- ometer, mount it and center it, and then go away. The device selects a good single crystal, orients it on the diffractometer, collects the diffractometer data, sends it over to the computing center which processes the data, makes the assumptions I had to make in my code, performs the symbolic addition, and what have you. You have to tell it in some way how much you're expecting, I guess. Maybe it can do that too, why not? Anyway, you end having all the information worked up automatically and the struc- ture poured out automatically at the other end. We haven't started on the paper-writing code yet. That seems to be all we need. Young: I thought one of the nicest things about AXI0M was the reason Bill Busing pointed out for the choice of its name: because you can take it for granted. I will structure my remarks by first stating what I now think to be the question we were addressing (for it is not the question I thought it was when we came in), by then giving some pros, cons, and alternatives, and by winding up with a recommendation. It seems now that the question may be stated as follows: With two demonstra- tion projects and a feasibility study now under way with NSF money, with the ARPA Network now in operation and about to go commercial, with all three agencies we've heard from expressing strong interest in networks involving fewer major computer centers, the question is not whether we think network and remote user systems with capabil- 122

OCR for page 119
ities suitable for crystallography should be established but, rather, whether we want crystallography to be counted in on developments that are obviously occurring anyway. The question might even be raised of possibly increased future difficulty of getting computer dollars from the agencies if crystallographers failed to take what the agencies think to be the advantage of these developments. We should at least look at them carefully. If we decide not to take advantage, we should certainly know why in a rational way. The crystallographic community has a strong history of coopera- tive effort, both national and international. For example, the data compilations in the crystallographic field are widely recognized as outstanding examples of cooperative effort. So crystallographers would be a good group to consider cooperative efforts in computing, certainly. It also rather seems to me that this group is in fact in- terested and rather positively inclined toward doing so. Let us consider this question of remote users. I would like to summarize some of the main pros and cons I've heard stated at this symposium. On the pro side I have selected about five points. It seems clear to me that protein crystallographers need larger computers than are now available in any one place. Second, the possibility has been raised of a quantum jump resulting from almost a new di- mension of computing capacity that a network might make available (a network backed up, of course, by the most advanced computers and systems). Third, it appears that sufficiently reliable, high-speed data transmission is available, now or in the near future, to obviate the need for physical transport of data or results. I was much impressed with this possibility of 80 000 bits per second, and en- vious of the turn around time that Helen Herman quoted. My turn around time is nothing like that, especially not near the end of the quarter when all the students are crowding around the computer. Fourth, for the largest component of the cost of computing, which is refinement, most of us are perfectly happy to use somebody else's well-tested program, and do just that. In our laboratory we'll probably expand our use of other people's programs as systems such as Jim Stewart talked about become available and are adapted to our local computing conditions. Many of you are already using Stewart's X-RAY 70 or the Busing, Martin, Levy least-squares refinement program, to name but two of several widely used programs. As these new sys- tems get developed and checked out, it seems only simple wisdom for most of us to ride along on the work of these good programmers, and use their powerful, cost-effective systems. Presumably these good systems would then be available on a crystallographic computing network. Fifth, Ed Meyer mentioned the possibility of information retrieval from a central library, information retrieval not only 123

OCR for page 119
in hard copy but possibly on one's own display terminal. You'd call in to Chemical Abstracts or some center for Crystal Data Compilation (which is obviously going to get on a computer now), ask the necessary coordinates for a structure, and a stereo view would also appear on your display terminal. Finally, a number of other exciting possibilities were mentioned by Dr. Freer. On the con side, I heard the following problems raised. The first is the computing cost. In our campus environment, real costs are subsidized and hidden to some extent, and the amount of costs charged to the project is often a small part of the real cost. If we are ever faced with having to pay full real costs, we may find that we do not do as much computing. We cannot ride along indefin- ately on the institution's computing program. A second point against the network approach is that terminal and telephone costs are not trivial. Helen Berman quoted $917 per month for the terminal costs and about $200 a month for telephone service; to that one must add computer costs, for CPU time primarily. The sum of these charges gets to be fairly substantial. One has a fixed cost of $917 per month for the terminal whether one does any- thing or not. Clearly, one has to be in the crystallographic com- puting business deeply before it becomes worthwhile to meet the fixed costs for network use. A third point against, although I did not hear this said ex- plicitly, is the possibility of weakening local resources. While removal of the main crystallographic computing load from the local computing center might not be disastrous in itself, if the main com- puting load in several other fields were also transferred to a net- work, the reduction of support for and demand for services from the local center would certainly affect its development adversely. That would mean, ultimately, less computing capacity available to other faculty members and the professional staff. Furthermore, diver- sion of computing and its dollar support may ultimately produce a local cost problem that will reduce the amount of free time available to us and to our colleagues at our own institution. Another question I think was raised concerns our contribution to program development. An effect of going to a network might be to decrease the effort put into this area because of (1) the difference in the cost basis of the computer time available, and (2) the possible feeling of awk- wardness in treating a program remotely. The result could be less rigorous development of better computing programs and, even, hard- ware. A further point against whole-hearted committment to networks is the specter of disaster wreaked if, for political or economic reasons, the network failed or stopped rendering services. An interesting alternative was offered by Dr. Sparks, and Dr. Larson has addressed himself to it. It is the use of an "adorned" 124

OCR for page 119
or "festooned" minicomputer in association with the diffractometer. If it were decided that crystallographers should consider getting onto a network, it seems to me that Ed Meyer has raised a number of specific practical questions that have to be answered. A small committee might be formed to study these questions further, along with the possibilities Steve Freer has raised, and other points, to determine whether it really is worthwhile for crystallography, as such, to opt for network computing as a prime resource. It seems to me we have made a good start, and have a good basis for further informed study of the options. Larson: A further point I wish to bring up is that the development of this next generation of computers is along the line of parallel processors, which are essentially vector processors. If you think about the protein problem, or for that matter any crystallographic least-squares problem, the whole thing can be quite easily broken down into a vector procedure. Some of the problems we are contem- plating may make much better utilization of these machines than the vast majority of other types of projects or other types of number- crunching games that, for instance, I am competing with. We have long vectors; with a thousand data points a vector of hkt is a thousand points long. For a protein, you have a vector of x,y,z that is 3n long. These vectors are long enough so that these vector machines can start to move. I think the protein people should ser- iously address themselves to the possibility of using these machines. Hamilton: I had hoped to have someone here to tell us about the capa- bilities of ILLIAC-4. A related point is that once the protein peo- ple decide which really are the best algorithms for the refinement of their structures, one might even design special hardware to carry out these algorithms very specifically. Dewar: Once one has a network one no longer has a drive to move things around to all computers on the network and one can't afford to write highly specialized code for highly specialized hardware. I'm sure that if 10 million dollars was spent on crystallographic computing several million of it could have been saved by rewriting critical parts of least-squares programs in assembly language. I think there's no doubt about that, and it becomes a lot more practical and at- tractive on a network where you can do a lot of local tailoring and use it in situ in a place where it's going to work. Larson: This comes back to the question of developing programs. Over a period of many years at Los Alamos, from the time we started to do single crystal work up until three or four years ago, it was hard to talk us into even giving away a program. The major reason was that we felt that in general a person who knows what his computer 125

OCR for page 119
program does is more likely to use it correctly than someone who is using one taken on faith from someone else. So, there are many considerations to balance. Frey: For routine structure determination, first question, is it useful to refine to death? Second question (if not, as I think), is it useful to save the data you collected? Larson: No, it is not useful to refine something to death. No it is not useful to save the data you've collected. Frey: But I know no means, no way actually at the moment, to save, say, a factual structure determination. It's difficult to see what I'm going to publish. Hamilton: There are ways to save the data, as you know, by publishing them in Acta Crystallographica and various data banks being main- tained, so that if somebody else wants to come along later and refine them to death, perhaps they can use their computing time to do it. I think that's a good point—by all means collect the data as well as you can, but perhaps only abstract the information you need from it yourself and save it for someone else to carry on with. I think there are people who would argue the point. Larson: I'll argue with this point. I think you should collect the data to do what you want done. If you want to find the atoms of the struc- ture and you don't care about the thermal motion, collect the data in a mode that will locate you the atoms and give something that neither you nor anybody else will believe as the thermal parameters and then throw the data away when you get through with it. You might publish the numbers, but the cost of recollecting data on moderately small structures is actually very small. How much time does it take for an automated Picker diffractometer to collect 3000 data points? Our standard procedure runs about 600 a day. That's five days. And it runs while I do something else. Dewar: If it's desirable to maintain data banks, clearly networks and trillion-bits stores are useful. Larson: The task you're talking about is presumably protection of a local software system, and this is essentially the security problem we are up against. If you put a computer onto a system of this sort, you must have the system designed with safeguards such that each and every individual computer can detect any attempt by the system or network to steal information. There should be no way for computer A to find out anything about the system or data or anything else at computer B without permission. There should be no way to find out about it's system and the possibility of acquiring 126

OCR for page 119
data will require knowing sufficient passwords and other things to get through the red tape. Anonymous: You're saying there will actually not be perfectly free access from my local computer to any other in the net? Larson: Would you like Walter Hamilton to have free access to your carefully collected structure factors? Anonymous: I'm not talking about that. I m talking about such things as optimizing a PL-l compiler, for example. Dewar: There is a significant question here having to do with whether IBM will be upset when the number of PL-l compilers on order goes down by two orders of magnitude. I'm not quite sure what one can hypothesize about that situation. IBM will probably work out some- thing that seems reasonable to them and that seems reasonable to people who use the network, and they'll buy it. Otherwise they won't. Thomas: There is another implication of networking that Ray Young men- tioned but that hasn't been made explicit enough. We might all agree that there is developing a dependency on computers, not only by crystallographers but by everybody else, which approaches the dependency that we have on electric power for example. That kind of dependency can be catastrophic unless sufficient redundancy is built into the networks and systems. Thus a network must not be allowed to fail totally, for hardware, software, or almost any other reason. One of the points of this symposium should be to express our concern regarding the required reliability and redun- dancy that must be built into networks in order to avoid the kind of collapse one can easily foresee. Hamilton: That's a very good point. If all of the crystal structures are being refined on one computer that has one bit wrong in some position, then all the structures in 1972 will be wrong for the same reason. Dewar: We seem to be getting swept away by the ARPA Network and for many of us it's a first exposure. But there is another view which is that the computations we're performing are basically trivial. To a computer scientist, for instance, they are general, perhaps in the last analysis even boring, problems just consisting of simple arithmetic, and part of what we are fighting today is that big machines are not designed for simple arithmetic. They're designed for data retrieval, swapping, paging, devouring resources, etc. There is a cogent argument for establishing a big computer designed solely for numerical processing, getting rid of all these problems that come from information handling. From this point of view, elim- 127

OCR for page 119
inating completely extraneous uses of the computer, one can certainly get utilizations and efficiencies that are greatly enhanced, and I think we should not neglect the possibility that the most effective way of dealing with crystal structure solution may be to establish one crystallographic computer specifically oriented towards that task. The aims are similar enough to those of the quantum chemistry study that there is a confluence of objectives here. Sayre: A National Computation Center does not in itself satisfy the desire for rapid transmission of problems from smaller installations. Furthermore, the placing of such a Center in a network would not preclude the development of special-purpose hardware at that Center aimed at providing particularly efficient service for crystallogra- phers and other special groups. Therefore it seems to me that the committee considering such a National Computation Center might stipulate that such a Center, if it is to be created at all, should be created within the framework of a computer network. Dewar: I just want to point out that the existence of the network today may not be the entire solution. We may still want particularly to look at special-purpose hardware. Kay: If networks are going to be set up in the next few years, off- shore areas such as the Caribbean where both I and my colleagues such as Fletcher work should be remembered. Hawaii, Alaska, or even Montana, have similar problems. Our telephone system is rather poor and the international lines are probably not fast, and are very noisy. I wonder if satellite transmission has been thought about for remote areas? It would be helpful if some provision were made for interna- tional usage for the benefit of undeveloped areas. Interested peo- ple in Jamaica or Venezuela or other places that have moderately act- ive groups could probably make good use of U.S. data bases. It could aid in their economic and educational development. Lykos: There already exists a system in Hawaii called ALOHA. They have a version of the IMP called MENEHUNE ('little elf). What is dif- ferent about their system is that radio transmission is used rather than copper wire and microwire (i.e., telephone circuits). In fact, Larry Roberts, ARPA director of the ARPA Network, discusses the im- plications of radio transmission between a handheld transceiver and a computer/communications system at the 1972 Spring Joint Computer Conference, of which Proceedings will be distributed at the Conference (May 16-18, 1972). With regard to your second point, satellite data communication is already happening. In October of 1972 there will be an International Conference on Computers and Communications at Washington, B.C., designed to provide a forum for technologists in Computer Science, Economics, and Law. A demonstration project is planned involving India, England, and the USA. Should the world 128

OCR for page 119
of crystallography have a contribution to that demonstration, I would be happy to provide a contact. Oxman: The National Library of Medicine's Lister Hill Center for Bio- medical Communications is currently involved in trial use of a NASA satellite for transmission of voice and data communications between the center, located in Bethesda, the University of Alaska, and designated clinical centers to improve health-care services in Alaska. This is perhaps a basic research effort to develop and test the feasibility of using satellites for the communication of data. Fritichie: I'd like to ask the representatives of the government agencies, in view of their encouragement of large computing networks and cen- ters, what is the attitude towards scientific data banks, such as for example a depository for crystal structure information. A decade ago or so this sort of thing was discussed, but the technology then essentially was not up to the conception. Has that point been passed and are the agencies now willing to put money into support of con- struction of such data banks? We seem to be close to the point of eliminating publications. If everyone is hooked into a data cir- cuit and has access to data, once they have been appropriately re- ferred and deposited, why publish at all? Lykos: You are speaking directly to the programmatic thrust of the NSF Office of Scientific Information Systems. Program and data sharing by network not only provides a medium of communication between scien- tists with similar interests but also between scientists who are finding channels of communication via problem-solving algorithms im- plemented as computer programs. Of course this will have an impact on the very process of publication. Crystallographers who want to get into the creation, maintenance, and accessibility of data banks should get in touch with OSIS. Hall: Crystallographers should not get as involved in systems program- ming as we have done in the past. I refer particularly to machine and assembly language coding. If this is needed for special ap- plications, our efforts would be a lot better employed if we directed people who are really trained to do it. We all agree that if you replace the much-used program DO-loops with machine code you can speed things up, but with the optimizers available in most compilers today the gains are minimal. Much more time and money have been lost over the years in changing other people's programs from an assembly code back into a form that will suit your machine than has been gained by this operation. With vector computers of the future, again I think it should not be our concern to optimize the vectoring of our programs. Our objectives with a computer such as ILLIAC-IV should be to ensure that the systems analysists supply a compiler that will 129

OCR for page 119
vector it efficiently for us. The crystallographer at large should not have to worry about such systems problems. I would like to put in a plug for generalized FORTRAN programming that can be used in a black-box fashion. Allen Larson suggests that all users must un- derstand the workings of the crystallographic programs they are using. I think those days are well and truly past. Already we have chem- ists, mineralogists, and others using our program who know little about crystallography per se. Such use is likely to increase with time, and I do not view it as a bad thing. Dewar: A lot of people who are normally crystallographers are actual- ly first assistants to programmers. A lot of people working in crystallographic laboratories are really first-rate programmers who have contributed a tremendous amount of computer technology. One of the ideas back of the proposal for a National Center for Computa- tion in Chemistry is bringing together a group of highly trained personnel, not just hardware. Some of that tremendous expertise is terribly diffused at the moment. I cannot agree that it is not worth looking into specialized programming. There are people who have in- terests right up that street and, given ideal situations, could achieve tremendous things. Certain chemical crystallographers should not get involved, but there are others very much interested, and they should be given an environment in which they can work effectively. King: A crystallographer cannot be expected to design the system, but an appreciation of what the systems can do would avoid binds where some FORTRAN code does not effect the complete transfer of one word. If you appreciate the differences between systems, you can sometimes avoid costly situations. Also, such understanding can help you write your FORTRAN code so as to use whatever is most efficient for the particular system. I have found that, whenever I wanted to write a subroutine in machine code in order to get a tighter loop, a nice way to do this when using batch processing was to get the compiler list- ing from the FORTRAN code, and then remove the redundant steps. This went about as far toward making the loop tighter as anything one could go, and besides, one learned a little bit about the system, even if only about the practical details of the FORTRAN system that could be safely eliminated. Now that we are using a remote terminal, I regret that we do not get a compiler printout to tell us how to improve our own FORTRAN programs. Hall: One should not have to worry about whether a half word is trans- ferred in a computer operation, or not. The computer people should be made to fulfill the specifications as defined in ASCI-II, etc., and our efforts would be best used in ensuring that such standard- ization is adhered to. As users, we should not have to be concerned about such special hardware operations. Larson: I think it should not be necessary as you say for us really to 130

OCR for page 119
worry about it. The one who's doing the systems programming like Jim Stewart or myself has to worry about that. He has to have it done so that if we deliver these systems to you, you can get them on without too much trouble. It's sometimes very difficult to do that. Anonymous: I would like to echo Dr. Hall's sentiments on this matter. Professor Dewar started out saying that crystallographers should be concerned about the social importance of their problems, but then he turned around to chide us for not putting our programs into special- ized machine language at a specialized computing center. The em- phasis has been so much on efficiency, as opposed to the thought that the networks and the national computing centers would perhaps enable the protein crystallographer to formulate and actually solve his problems. I would hope that the study of a national center would be much more concerned with the scientific value than with the ef- ficiency that would result. Dewar: I do not think you can separate the two, There are problems with price tags attached to them, and increasing efficiency is the name of the game, because that is what puts more problems within reach. Young: Surely we should do whatever we do efficiently, but we should not be so concerned just with efficiency that we fail to look about to see what else we could be doing. Steve Freer touched on that point when he suggested that the availability of this network may produce a new dimension in computing, a quantum jump. We may be able to take on new problems we had not conceived of before. Stewart: But the quantum jump will not come in crystallography if all of us are continually trying to second-guess the systems people on the network. Larson: You are, of course, pounding at lack of communication with the software designers. We are not apparently getting to them the mes- sage that FORTRAN-IV should be FORTRAN-IV or PL-l or whatever lan- guage . Young: That message is clear enough. Larson: But they are not reading it! We should not have to push that message to them. Those of us who are programmers should be able to take a FORTRAN compiler on any machine and presume that the FORTRAN deck will produce the same results on an IBM 360, on a Univac ll08, on a CDC, Honeywell, or GE. That is the reason you have gone to your pidgin FORTRAN, and even your pidgin FORTRAN is not legal on all machines. The biggest trouble I think you are having there on that half word is in the ability to use half-word integers or two byte integers. 131

OCR for page 119
Hamilton: Is there a consensus here that the crystallographic community is interested in exploring seriously the possibilities of crystallo- graphic computing networks and centers, and that this message should be carried to the group that is studying the theoretical chemistry center? It seems to me it would be a good idea if the ACA Computing Committee also were to discuss this matter. Larson: The ACA Computing Committee plan in their report to make a strong suggestion that people submit programs to the Quantum Chemistry Program Exchange. Hamilton: I was suggesting that the ACA Committee might give a few hours of good discussion to the topics we've raised today, perhaps to see whether the ACA wishes to stimulate any positive action on the part of the funding agencies. Any final comment? Fritchie: One problem that was mentioned fairly early has been somewhat neglected in our discussion. That is the human problem of maintain- ing a service attitude in such a center. The advantage of competi- tion from several equivalent centers has enormous potential, and I think this point should be considered carefully by any study group. Calvert: I would not like a consensus to go forward on the basis of one particular area of interest. As a society we should be care- ful to remember that a large number of crystallographers are in- terested in small or medium-sized molecules. The special problems peculiar to very large molecules might better be discussed sep- arately. It is possible that small structures can continue to be done simply and cheaply by existing techniques. Hamilton: I would agree with that. I certainly wish to thank all of the people who have participated, the principal speakers and members of the audience as well. The symposium is adjourned. 132