Technical Computing Alternatives: Supercomputers to Ordinary ...

'Ww3-

ISSUES & OUTLOOK

Technical Computing Alternatives: Supercomputers to Ordinary Computers

A I

By Gordon Bell Corporate Vice President and Chief Scientist Stardent Computer Inc.

ccording to Congress and the press, technical computing is in trouble. But in fact, the future looks brighter than ever. New classes of computers and new software are being created by new and existing companies. Only the growth rate for the traditional supercomputer might be slow. The reason is straightforward.A user can often do the same computation on a powerful PC, on a workstation, on a minicomputer or microcomputer, on a superminicomputer, a graphics supercomputer or 3-D workstation, a minisupercomputer, a special-purpose supercomputer, or even a traditional supercomputer. Users can trade execution time for cost because "Flops is Flops." The computer power necessary for technical computation can be substituted across this entire spectrum. Technical computing is moving away from highly centralized, time-shared supercomputers. The same forces that operated in traditional computing will pre vide distributed, interactive computers to technical users. These computers offer adequate capacity and peak power for demanding jobs, are cheaper and easier to purchase and use, and offer superlative price/performance ratios. A genembpurpose supercomputer is a machine that, at the time of its announcement, costs more than other computers (perhaps $5 tw $20 million), runs faster in general, has greater primary and secondary memory, and is suitable for all scientific, numeric problems. Supercomputers have evolved along one architectural path. They all employ vectors and powerful multiple processors to gain speed. Fortran is by far the most importantprogramming language;dusty deck programs are expected to port easily. Automatic compiler tools (vectorizers and such) help users exploit the new machines. The more adventurous eventuallyreprogram using explicitlyparallel techniques. Within a decade, Fortran will undoubtedly have parallel constructs.

Table 1: Power of 1989 Technical Computers in Megaflops/Second L m #Proc. per Max Proc.

ILFK per Ijnpack Machine lOW00

PC Workstation

MicroJMini Supermini Graphics Super Minisu per MainJVectors Supermmputer

1

1 1 6

1

0.1 to 0.5 0.1 to 0.5 0.2 to 1.5 0.5 to 3.0 0.1 to 0.5 0.1 to 0.5 4 1

4 8 6

1.5 to 5 2 to 4.3 7.2

10 43

8

19

150

10

6to 12 6 to 16 13 84

Peak

~

0.1 to 1.0

6 0.1 to 0.5 6

80 ,166 518

2,144

8 2 24

128

200 798 2667 .

22

...

... . ..

ISSUES & OUTLOOK

Applicationsmust be reprogram med to exploit these machines fully, and such programs can run faster than on a super. But the need for reprogramming and slow scalar speeds causes these machines to fail t h e workload test for serial job sbeams.They are not directly substi-

tutable for current, general-purpose supercomputers,

Super wtwlti~cessors(developed by BBN and Evans & Sutherland) have 100 or more processors, a common memory and a single job pool controlled by one operating system. The sheer number of processors guarantees good throughput performance, but the relatively slow individual processors will find it difficult to run scalar programs at supercomputer speeds. Still, these machines might be the highest capacity, general-purpose, cost-effective sys terns that can be built today. Ordinary computers provide the

mostkhnidcompuhgpwertday This includes PCs, which are evolving toward the power of 1-D, ZD and 2Y-Dworkshtionq %D high-pwhrm-

ance workstations; microcomputers;

programs after considerable tuning.

and superminicomputers. Many of

Notice that peak rates (some

these have impressive scalar performance, but they have no way to hit performance peaks for those programs that make good use of the vector or parallel capabilities of supercomputers. Table 1 summarizes the power of various technical computers in 1989. The machines in the bottom half are capable of providing shared supercomputer power. The column labelled #hoc. Max describes the parallelism available. The columns headed LFK.giveperformanceon the Livermore Forban Kernels, a good synthetic mixed workload for scientific machines. These numbers measurethethroughputone might obhin in a typical scientif~cenvironment The columns headed Linpack measure performance on the h p a c k linear equation solving test. The 100x100 test shows the rates that might be achieved by normal users of reasonable supercomputer programs,and the 1oooX1OOOtest shows rates that can be achieved by real

times listed at about 20 gigaflops) are not shown for the very large machines because no real progrdms come anywhere near the peaks. Hy significant tuning to run in parallel, programs operate at over onehalf the peak.

WHICH COMPUTER IS THE BEST FOR THE APPLKATIQ N? Market substi tu tion occurs across all computers. Users have a fixed budget to trade off across the complete range. Computer choice depends on many factors besides purchase price and peak perform-

ance: software availability, ease of purchase, installation and use; apparent lifetime; rate of technological change; past and future compatibility; control in the allocation and management of resources; programming knowledge needed; even the machine appearance or the prestige

Table 2 Installed Capacity for Technical Computing (Dataquest) Companies '89 LFK

Dataquest

Installed

Ships

Capacity

PC Workstation Micro/Mini Supermini

3.4M 0.4M 0.9M 0.3M

1M

100s

290K 51K 7.5K

1341 580

-20

Graphics Super Minisuper Parallel Roc. Mainflectom Supercomputer

10.5K 1.6K 365 8.3K

13.6K 600 250 1600

30 100 182

32

5 24

450

130

4 46

100

Selling Building Dead

7 7 2

3 4

-2

? ?

? -50

? ? 2 >2 >9 ? >3

-100 -10 2 8 8

3 3

-. 431

fSSUES & OUTLOOK

.....

. . . . . . . . . . . . . . . . . . . . . .,

r IwtiiiiK :I

~ ~ : ~ I ~ I cvoiiilw!(*t. wl:~r

Table 2 shows the installed capacity for techniral crrmpntinR, n e iriiisl iiiillorlmil r-oliiiiiii is 'H!)I.I;IC Capacity - the power available to run general-purposetechnical work-

.....

~ 4 ~ l l ~ ~ r :orl l i4rip*rr'otllpulCrp, ~~li alltl

iiwritig, clitmidry

the last generation will csnhue to

try, fluid dynamics and

:itid

bioctwiih medical im

widen. Howwrr, altrmatiw cornput- aEt- prnmsinn an- h i n E hnsfnrmrcl CTS tllilt ; i r * t b f i ~ t rnough will conlinue by t l i c twwly availabltb rlistributtd the bend to distributed computing power. A similar change occurred even for applications that were about a decade ago when works& tions were introduced into the deloads measured in units of CRAY Y- previouslys~~eYd~plications. Architecturally, the reasons are sign of digital systems and chips. MP/&. Most of the power is p m All the supercomputers and alvidcd by machines that are not super easy to see. Supercomputers require anythings; even much of the super- expensive, high-speed components. ternates are highly parallel - using computerpower is provided by wimp elaborateprocessor-memoryconnec- vector prwessors, parallel CPUs or machines at least one generation old. tions, very fast, large disks, process multiple processing elements. ComComputation on the m n g s y ~ , ing circuits that do relatively few puter scientists and applications tern is costly and inefficient. There i s operations per chip and per watt, specialists must cooperate to undera supercomputer at onr national extensive installation and hiEh owr- stand and makt. ziw of pamllrl comI: itm-ntcwy iii:ik iiiK :a n h Iivt .l y I i.ivi;il iitirIE costs; worst', Ilwy Ii;ivc litltc: putinx: itistiluticais ~ r i u s trt!sprd an (1 calculation for contractors Ihrough- architectural scalability. encourage research and teaching in out the United States. The supercomSupercomputer buyers must these areas. puter produces a picture.compresses have great needs, great dedication to The arts and sciences of visualithe data, sends it over a slow but ex- the support of the machine and great zation must be propagated into the pensive network. and graphics work- budgets. For almost all users, KO- technical computing community.The stations recompute the image for nomics inevitably dictate the pur- cost in new software will be more static display. The computation and chase of smaller systems connected than repaid by the value of new indisplay could be done on a powerful in networks. sights. If the technical cumpiiting workstation in roughly the same In addition to the pure econom- community can meet these needs, elapsed time without the network or ics, current supercomputers lack the our future is bright. . . . . . . . . . . . -. . . . - . . . . ..- .......... snpcrcomputer. The computer exists visualization capability and interac- . to support a super bureaucracy. tivity found in distributed computing. Let us compare an Ardent Titan Networks coupled with workstations III/2 (two processon} with a CRAY are not adequate to provide the same Y-MP processor. Titan JII was incapability. For example, the use of duced one year after the CRAY Y-MP spread sheets,drawing programsand and delivers about 9 megaflops LFK even word pmessing is qualitatively throughput. The Titan's throughput different using terminals connected isabout 1J2 that of a Y-MPprocessor to time-shared computers through for LFK, its 100x100 Linpack rate is LANs, in comparison with the use of about 1/6 that of a Y-MP processor; pmonal computers. and its peak performance is about I / 50 that of an eight-processorY-MI? BASIC SHIFT TO A Titan IIV2 costs about 1/20 INTERACTIVE AND what a one-processor CRAY Y-MP DlSTRlBUTED COMPUTING does. Furthermore, the Titan 111's speed doubled (and itspricedropped) fromitspredecessorinonly18months. A significant change in computThe shortest conceivablegestation for ing styles is occurring in providing a supercomputer i s three years, and truly interactive design and analysis Ciw to CAKE. 95% of every ddlar wc' n'ueivu g(w\ t o Iwlp five* lo sewn ycars is mort* likcly. using high-pcrformance. supcr workimpwetished people in the third stations and ki dphics supercomputworld. ers. These provide more than 10 Sitiall wonder t h t we're the SUPERS' GROWTH percent of a CRAY Y-MP processor k s t run,best nranaged chariry RATE DECLINE LIKE THE in America. capacity and peak power, are often MINI AND MAINFRAME? three to seven times cheaper per delivered megaflop, are purchased Supercomputerswill continue to and installed by a single user or small evolve more rapidly than minicom- group, and overcome many of the puters and mainframes and the per- disadvantages of supercomputers. formance gap between the latest Mechanical and petroleum end-

HaW

To Akke The

Dollar

smgm overseas.

WILL

.

....-.I

. . . . . .

24

- ............................

....

--

Sup.reomputingRsViar

..............

D m b rV m

....

........................

...