Two His­tory Lessons

29.Jul.2009 • by W.R. Wing

Last week I promised to con­tinue talk­ing about the Private-Sector meet­ing hosted at NCSA, but as I started col­lect­ing my thoughts, I real­ized that I will prob­a­bly be talk­ing about this for the next sev­eral weeks, and that I needed to pro­vide at least some back­ground and con­text so that peo­ple can appre­ci­ate just how dra­matic the recent advances in supercomputer-based mod­el­ing, sim­u­la­tion, and analy­sis have been. Broadly speak­ing, most com­mer­cial com­pu­ta­tional mod­el­ing falls into three cat­e­gories: Com­pu­ta­tional Fluid Dynam­ics (CFD), Finite Ele­ment Analy­sis (FEM), and Com­pu­ta­tional Chem­istry. The first two cat­e­gories got their start in the mid-1960’s when main­frame com­put­ers finally became pow­er­ful enough to start to address sim­pli­fied ver­sions of inter­est­ing prob­lems. Com­pu­ta­tional chem­istry was actu­ally started ear­lier, but the first effi­cient ab ini­tio meth­ods didn’t appear until the mid-1970’s.

One of the first real suc­cess sto­ries for CFD was the Boe­ing 737. Boe­ing had been study­ing the idea of a small, (less than 100 pas­sen­gers) air­plane that would be self-servicing, and could fly in and out of air­ports too small to have ter­mi­nals equipped with jet-ways. Since Boe­ing was late to the game, the engi­neers decided to use the fuse­lage struc­ture from the 707, which not only saved time, it allowed 6-abreast seat­ing (a plus for air­line adver­tis­ing – the inside didn’t look so small). This cre­ated a fairly short wide fuse­lage, which would have made rear-mounted engines both inef­fi­cient and sources of stress on the fuse­lage. Mount­ing the engines under the wings solved this prob­lem, and in addi­tion put their weight close to the cen­ter of lift, which reduced stress even fur­ther. How­ever, in order to make the plane self-servicing, the fuse­lage had to be close enough to the ground to allow for inter­nally stowed board­ing stairs. This in turn meant that the engines couldn’t be pylon-mounted below the wings. Instead the engine mounts were made a for­ward exten­sion of the air­foil and the tops of the engine nacelles blended into the wings. The wind tun­nel tests were dis­as­trous. The wing area above and to either side of the blended engine mount wasn’t pro­vid­ing any lift and was pro­duc­ing fero­cious amounts of drag. Extend­ing the wingspan could fix the miss­ing lift, but the drag was so bad as to call into ques­tion whether the plane could ever fly profitably.

It was at this point that water cooler con­ver­sa­tion between the wind tun­nel engi­neers and the CFD mod­el­ers sparked inter­est on the part of the mod­el­ers. Because of the exten­sive wind tun­nel data, it was a per­fect oppor­tu­nity for the CFD engi­neers to refine their mod­els and con­firm their accu­racy. In the end, the answer turned out to be fairly pro­saic; it was that the wind tun­nel mod­els had been built using the same approach Boe­ing had used for all its other jet planes with pylon-mounted engines. The engines were mocked-up as non-breathing solid slugs. This worked fine for engines at the end of long pylons, but didn’t work at all for engines that were, in effect, exten­sions of the lead­ing edge air­foil. The answer, which was sim­ply to replace the solid engine mod­els with ones that were aspi­rated, was diag­nosed by CFD mod­el­ing, and later con­firmed by wind tun­nel tests. The 737 went on to become the sin­gle most pop­u­lar com­mer­cial air­craft ever built.

An anal­o­gous story exists for FEA. In late 1970s, the Ford motor com­pany was start­ing devel­op­ment of the Tau­rus sedan, and for the first time decided to use a finite ele­ment model to sim­u­late the results of crash tests. The codes had become good enough at that point that they could repro­duce defor­ma­tions of even fairly com­plex shapes fold-for-fold, crease-for-crease and bend-for-bend as long as the mate­r­ial in the indi­vid­ual parts was homo­ge­neous. The idea was not to sup­plant the USDOT-mandated crash tests, but to fig­ure out through a process of iter­a­tive refine­ment how to sim­plify the design, make it cheaper to man­u­fac­ture, and still pass the crash test. That process was not only suc­cess­ful, it was so suc­cess­ful that GM imme­di­ately insti­tuted a sim­i­lar program.

So, what’s the point? The point is that these water­shed events in com­mer­cial mod­el­ing and sim­u­la­tion were car­ried out on com­put­ers that, at best, were capa­ble of 160 mil­lion float­ing point oper­a­tions per sec­ond (160 MFLOPS) and had a max­i­mum of 8 megabytes of main mem­ory. Twelve years later, in the early 1990’s, the Intel Paragon super­com­puter was being deliv­ered with 1000 times this level of power (160 GigaFLOPS), and twelve years after that, Cray was deliv­er­ing Cray XT3 com­put­ers with 1000 times this again (160 TeraFLOPS).

What can you do with one mil­lion times more com­put­ing capa­bil­ity than was used to crash test the Ford Tau­rus or model the aero­dy­nam­ics of the Boe­ing 737? You can do amaz­ing things. Finite Ele­ment Mod­el­ing is no longer lim­ited to com­po­nents that made from homo­ge­neous mate­ri­als. They can be lam­i­nated, fiber-reinforced, and even include mate­ri­als that change phase or other behav­ior under stress or shock. This has led to major advances in crash wor­thi­ness, and weight reduc­tion, to name a sin­gle industry.

Com­pu­ta­tional Fluid Dynam­ics has moved from incom­press­ible, invis­cid (no vis­cos­ity, no fric­tion), “panel” mod­el­ing to com­press­ible, tur­bu­lent, dis­si­pa­tive flow. More impor­tantly, it has started merg­ing with com­pu­ta­tional chem­istry to model processes like com­bus­tion where it fol­lows not only the fluid flow, but also the chem­i­cal reac­tions and reac­tion prod­ucts. This is allow­ing mod­el­ing of every­thing from jet engine com­bus­tor cans, to diesel engines. In both cases, these are areas where even one per­cent improve­ments in effi­ciency have major eco­nomic consequences.

Tags: HPC, industry, NCSA