Involvement of Industry in the National High Performance Involvement of Industry in the National High Performance Computing and Communication Enterprise Computing and Communication Enterprise

We discuss aspects of a national computer science agenda for High Performance Computing and Communications (HPCC). We agree with the general direction and emphasis of the current program. In particular, the strong experimental component and linkage of applications with computer science should be continued. We recommendaccelerating the emphasis on \nationalchal-lenges" with more applications and technologies from the information, as compared to simulation areas. We suggest modifying the grand challenge concept to complement the current teaming of particular computer science and applications researchers. We would emphasize better linking of each application group to the entire (inter) national computer science activity. We express this in terms of a virtual corporation metaphor. The same approach can be used to involve industry in HPCC for both the consumers of HPCC technology (application industries) and producers|Independent Software Ven-dors (ISV) and the hardware system companies. We illustrate this approach with InfoMall, a HPCC technology transfer program funded by New York State. The federal program should have greater incentives for the involvement of both ISV's and their products.


Introduction
This commentary on the National High Performance Computing and Communications Initiative, and the associated National Information Infrastructure does not set forth a complete new program.This is not necessary|the current federal program is \almost right on."Rather we present a set of remarks based on the in-volvement of applications and industry in the program.We will consider both the industries that use HPCC as well as those that (will) produce the new HPCC software and system products that will drive the integration of HPCC into the mainline computer industry.

Opportunities for HPCC in Industry
While at Caltech, one of my major activities involved studying the parallel software and algorithms needed by large scale scienti c and engineering computation Angus:90a], Fox:88a], Fox:94a].One of my interests in moving to Syracuse was to extend this approach, but focus on industrial and not academic applications.This activity, now called InfoMall Fox:93c], Mills:93a], Mills:94a] is supported by New York State and aims to accelerate the introduction of High Performance Computing and Communication technology into the State's industry.It has three major components.Firstly, it involves those industries that uses computers (i.e., the applications of HPCC, which are described later in this section).Secondly, InfoMall seeks to create (enhance) an HPCC software industry in the State, and Section 3 describes issues connected with this activity.The nal component of InfoMall is devoted to training and is especially centered in the mid-Hudson (Kingston, Poughkeepsie, and Fishkill) area of New York where some 10,000 former IBM engineers are looking for new job opportunities|InfoMall aids those interested in HPCC careers.Such pools of engineers exist in many parts of the country, a byproduct of the major restructuring in aerospace, manufacturing, computer, and other industries.This national resource indicates that we have the people to implement the integration of HPCC into the country's industry.We \only" need to identify the viable opportunities and the necessary funds (venture capital) to implement them.Now, we discuss which opportunities are in fact viable|this returns to the rst component of InfoMall|what are the relatively near term industrial applications of HPCC.The rst phase of our New York State project involved a signi cant survey reported in Tables 1 to 6, which divides possible applications into four major areas, and then in more re ned fashion, 33 classes.The good news is that as summarized in Table 2 and Column 2 of Tables 1 to 4, we already have the necessary, if not the most productive, software technology needed to implement all of these applications.Further, extensive computer science research tested in grand challenge and other academic and research implementations, has given us excellent parallel algorithms.Again, a good argument can be made in every entry in Tables 1 to 5, that HPCC will make a di erence and improve the performance and/or competitiveness of the associated industry.Nevertheless, only in a few cases will the return on investment justify the work involved in implementingthe HPCC application.This conclusion is explained in Fox:92e] Mills:94b], Fox:94b], Fox:94c], and here we can only summarize and illustrate this conclusions.Note that the four application classes in Table 1 are: (A)    Originally, I had envisioned a major InfoMall activity in the Simulation (A), as this was the natural counterpart to my personal academic activity at Caltech, as well the broad national endeavor in the Federal Grand Challenge Initiative.However, the clear near-term value of HPCC simulations in the research applications is not so obvious in industry.We can only illustrate this conclusion with one example here.Consider the application of HPCC to manufacturing|an area of exceptional importance and promise.One can use HPCC for one or more of the components of manufacturing|such as the rst three classes of Table 1| uid ow, structural and electromagnetic simulation.However, these are in general quite small parts of the whole manufacturing and design process.Thus, we see Amdahl's law applied to manufacturing|if simulation is (for example) 10% of the design cycle, then parallelizing this one part can give at best a speedup of 1.1 We are still dominated by the 90% remaining \sequential" part of the process to which we have not applied HPCC.Thus, the best opportunity for HPCC lies in its integration into the  entire manufacturing operations.This will enable agile manufacturing and as components of this, concurrent engineering and multidisciplinary analysis and design.Just considering the latter, this is as shown in Figure 1 a challenging but in principle, perfectly feasible integration task.However, the real world makes it \very hard".Thus, components of Figure 1 are existing one million codes (such as NASTRAN) which not only have to be parallelized, but also integrated with other major software packages.Perhaps more important di culties are nontechnical issues.Most manufacturing companies face restructuring, cutbacks and/or erce global competition.Agile manufacturing implies not only technology investment and (risky) development, but remaking the whole engineering process.Thus, we cannot address manufacturing just as an HPCC technology problem, but many national infrastructure and education issues are involved.I see this area as an exciting long-term area for HPCC, which need major government investment as industry will nd it hard to justify the investment based on a typical one to three year return.
Returning to the survey of Tables 1 to 6, we found (in New York State) much more near-term promise in the information related applications (B), (C), and (D).There are several reasons for this.The simplest reason is phase space|information processing always has been the dominant use of computers in business and this is likely to be accentuated and not decreased in the future.Further, one \only" needs to parallelize a few software pieces (e.g., implement a parallel database) to enable a large set of applications.Although very chal-lenging, this task is well underway commercially with, for instance, Oracle 7.1 supporting parallel query.In contrast, simulation has no single \killer application" and each piece of software is typically only used in specialized fashion.The National Information Infrastructure (NII) applications in (C), parallel servers on highspeed networks, have another advantageous feature| namely they are \new", that is, involve developing software from scratch and not as in many simulation applications, porting existing sequential codes.It is clearly easier to justify the investment in parallel machines for new software systems than for the cases when porting of large existing systems is involved.This is a rambling story and incomplete to boot| what do we deduce from it?I concluded that information processing is the most promising role for the HPCC industry.Correspondingly, I have redirected our New York State funded activities in this area.Our project must focus on relatively near term opportunities to satisfy its (funded) economic development goals of job creation.However, the federal government needs to support both simulation and information processing related technologies|HPCC simulation will be critical in the long run and not an easy investment for industry.Thus, I support the broadening of the Grand Challenges to include a new set of National Challenges focussed on information based applications.This should not only feature NII applications (C) (and parts of (D)), but also the basic information analysis cases in Table 3.In fact, as the NII does not yet exist except as the comparatively low bandwidth internet, industrial interest will initially center on those cases in Tables 3 and 5, which do not require the pervasive multimegabit/second network for everybody, everywhere and at every time promised by the NII.

Technology Transfer and the Role of Grand Challenge Teams
The approximately 50 Grand Challenge teams have been correctly set up as collaborations between computer scientists and application types.This can help the particular computer scientists if the application uses, evaluates and motivates further development of their personal technology.I have seen this is in many grand challenges and have concerns because this is important, but only half the story.Thus, we also need to understand and develop the best computer technologies in the world for this application.From this point of view, we have the application and not computer science driving the technology.This identi es another role for the computer science members of the grand challenge team.They should act as a window (educational/knowledge source) into the best HPCC technology that one can bring to bear on the application.One needs applications to act as testbeds for new HPCC technologies.However, in the grand challenges, I suggest that an application pull is most appropriate, and that in general other (funding) mechanisms be used for the application testbeds (computer science pull) which are also denitely needed, especially at the initial stages of technology development.
We have embodied this concept into the model for technology transfer in InfoMall, which is illustrated from two points of view in Figures 2 and 3. We have a world wide customer (application) pull (on right side of gures) that feeds from a world wide pool of HPCC technologies (left side of gures).These are fed through the \window" discussed above into a software capitalization (InfoTech) process that produces usable prototype systems, which can be commercialized at lower risk than raw technologies.Key to this InfoMall engine is the entrepreneur or ISV (Independent Software Vendor) who takes InfoTech products, and with traditional economic development support, produces the HPCC products brought to the customer by the marketing and system integration components of InfoMall.InfoMall o ers to ISV's the special high technology support (Table 8) needed for an entrepreneur in HPCC.This includes both access to facilities and educational resources, as mentioned earlier in context of our mid-Hudson project.
We believe that major e orts should be made to increase the role of all parts of industry in the HPCC program.This should include the users with realistic estimates of the time scale on which HPCC will be useful technology|as illustrated in Section 2, this will certainly vary from case to case.However, the ISV's are also critical for there can be no HPCC hardware or systems business, and no signi cant user base, without ISV's linking them.We see at least three ways how the HPCC program could encourage the HPCC software industry.
1. Encourage federal proposals that use commercial rather than home grown software systems.I am reminded of a story of one computer vendor who complained that computer scientists could not buy their system because the commercial systems software was too good|one could only get federal funding to build software for systems where the existing commercial o erings were poor.Instead, we should encourage research that builds on top of existing commercial software.We should encourage the latter to o er open interfaces enabling this strategy.2. Change the funding of MPP to explicitly encourage purchase of software, as well as hardware|recently, I was told that HPCC systems software companies were nonviable.One couldn't sell HPCC software  We need open (free) interfaces and not necessarily free software.The above picture is that of a virtual corporation with products built of interoperable multiuse modules|each module in principle built by a separate organization, which are linked together in InfoMall to give a complete product.The HPCC national endeavor o ers the possibility of a far larger and more exciting and important virtual corporation.This will only be possible when greater encouragement is given for software and user industries to participate in and bene t from HPCC.

Figure 1 :
Figure 1: Software Structure for Multidisciplinary Analysis and Design

Table 3 :
Data or Information Analysis (C)Table 4: Access to and Dissemination of Informa- tion (D)

Table 5 :
Integrated Information Services including

Table 1 :
Simulation Opportunities for HPCC in Industry

Table 2 :
Simulation Opportunities for HPCC in Industry (contd.)

Table 7 :
Runtime Software Support for Tables1 to 5