Exploiting Market Realities to Address National Security’s High-Performance Computing Needs Mark D. Hill Computer Sciences Department University of Wisconsin--Madison (Modified 01/00 for IDA-Bowie) IDA Defense Science Study Group 11/99 Mark D. Hill, University of Wisconsin-Madison Bottom Line • Recommendations – – – – Avoid directed procurement Exploit clusters Provide sustained funding to academia Build the high-performance computing (HPC) market • Talk Outline – – – – – HPC is important to national security History: PVPs and MPPs Future: Clusters Four Recommendations Summary IDA Defense Science Study Group 11/99 Mark D. Hill, University of Wisconsin-Madison HPC is Important to National Security • HPC is the upper extreme of computing – Supercomputers, etc. , costing $10M-$100M • HPC delivers for some important problems – Breaking encoded messages (NSA) – Nuclear stockpile stewardship (e.g., LANL) (without nuclear testing) • Requirements – Trillions of operations per second (tera ops) – Trillions of characters of semiconductor memory (terabytes) – 1015 characters on disks & tapes (petabytes) IDA Defense Science Study Group 11/99 Mark D. Hill, University of Wisconsin-Madison History: Parallel Vector Processors (PVPs) • E.g., 1980s supercomputers from Cray Research – custom processors (i.e., no microprocessor) • Assessment – Were ideal for NSA and LANL in 1980s – Sales hurt for “killer micros” – E.g., Silicon Graphics Inc. (SGI) buys Cray in 1995 IDA Defense Science Study Group 11/99 Cray-1 Mark D. Hill, University of Wisconsin-Madison History: Massively Parallel Processors (MPPs) • E.g., Cray T3D, Intel Paragon, & Thinking Machines CM-5 in early 1990s – Replicated identical hardware (especially microprocessors) – But specialized software – Integrated computer vendor • Assessment TMC CM-5 – High-end HPC market stalls (see next slide) – Big companies lose interest in HPC – Small companies go out of business IDA Defense Science Study Group 11/99 Mark D. Hill, University of Wisconsin-Madison HPC Market in Billions US$ [IDC99] • 1.8 1.6 1.4 1.2 1 0.8 < 1M ASP >= 1M ASP 0.6 0.4 0.2 0 1993 1998 IDA Defense Science Study Group 11/99 Mark D. Hill, University of Wisconsin-Madison Future: Clusters • Clusters of “Nodes” – Nodes: PCs to commercial servers – Networks: connect nodes with standard network to custom “system area network” – Cluster software: optional software that makes cluster appear more like an MPP – Clusters are a big part of DOE’s Advanced Strategic Computing Initiative (ASCI) IDA Defense Science Study Group 11/99 Mark D. Hill, University of Wisconsin-Madison Future: Clusters, continued • Nodes – Commercially viable • PC or server nodes? (You get what you pay for) – Available from multiple vendors – Insures sustained availability • Network – – – – Commodity LAN? Specialized SAN? You get what you pay for Need to port to new LAN/SAN every few years • Use middleware to rise above the details (e.g., MPI) IDA Defense Science Study Group 11/99 Mark D. Hill, University of Wisconsin-Madison Future: Clusters, continued • Software – Need to port to new cluster every few years – Think about the performance lost from software not working in the first 1.5 years of a cluster’s 3-year life • Use shrink-wrapped software whenever possible • Use standard languages & libraries • For custom software: K.I.S.S. • Integration -- Creating a “Computer System” – Hardest problem (done by MPP computer vendor) – Select & deploy network hardware, network protocols, middleware, application library, debuggers, etc. – Who does integration? Customer? 3rd party? IDA Defense Science Study Group 11/99 Mark D. Hill, University of Wisconsin-Madison Recommendation 1: Forget “Procurement” • Can’t buy HPC like bombers or carriers – Computer technology moves too fast for contractor mentality (two times the performance in two years) • Time lags too long • Specifications too detailed – Can’t depend on sustained government commitment (post cold war) – Can’t depend on one company (or a few) • large - “zero billion dollar market” • small - out of business IDA Defense Science Study Group 11/99 Mark D. Hill, University of Wisconsin-Madison Recommendation 2: Exploit Clusters • Government should buy Clusters – Customer (e.g., NSA or LANL) • Is responsible for mission • Should be responsible for cluster integration – But • Sub-contracting integration? • Avoiding duplication of effort? – Many technical problems • Clusters are like Churchill view of democracy • “Democracy is the worst form of government except all the others that have been tried.” IDA Defense Science Study Group 11/99 Mark D. Hill, University of Wisconsin-Madison Recommendaton 3: Fund Academia • (Warning: I have a bias here) • Sustained funding of academia to develop new HPC ideas (as we have done in the past) – Not just Kuhnian paradigm shifts (DARPA) – Not just “trickle down” (DOE ASCI) – The country reaps what it sows IDA Defense Science Study Group 11/99 Mark D. Hill, University of Wisconsin-Madison Recommendation 4: Build HPC Market • Government should encourage private HPC customers to reduce mismatch between needs of government and private sector – Demonstration projects – Personnel exchanges – High risk but high payoff – Unbounded potential in Biology IDA Defense Science Study Group 11/99 Mark D. Hill, University of Wisconsin-Madison Bottom Line • Recommendations – – – – – Avoid directed procurement Exploit clusters Provide sustained funding to academia Build the high-performance computing (HPC) market 10 issues in white paper • Acknowledgements – DSSG: Gould, Licato, Major, Roberts, colleagues, & mentors – IDA or CCS: Brenner, Carlson, Draper, Feustal, Greenberg, & Mayfield – LANL: Cerutti, Lee, Luo, McCoy, Thompson, Reynders, Wasseman, Watson, & White – NSA: Powers / DARPA: DSSG sponsorship & Hendler – NSF, Compaq, IBM, Intel, & Sun: my Wisconsin research sponsorship IDA Defense Science Study Group 11/99 Mark D. Hill, University of Wisconsin-Madison Complete Recommendations (1 of 2) • Don’t direct one or a few computer vendors to build PVPs or MPPs (like bombers or carriers) • Do provide money for government to buy HPC machines (like ASCI but it’s not sufficient) • Encourage HPC customers to take responsibility for cluster integration • Appealing to industry patriotism will not work • Provide some money to computing vendors to build better HPC machines IDA Defense Science Study Group 11/99 Mark D. Hill, University of Wisconsin-Madison Complete Recommendations (2 of 2) • Expose industry to HPC potential (pilot projects & personnel exchanges; consider biology) • Expose academia to HPC potential (pilot projects & personnel exchanges; consider biology) • Provide sustained HPC funding to academia • Provide HPC benchmarks to academia • Provide some money to academia to buy HPC machines to build better HPC machines IDA Defense Science Study Group 11/99 Mark D. Hill, University of Wisconsin-Madison