krj@utcsri.UUCP (08/24/87)
NA Digest Friday, August 21, 1987 Volume 87 : Issue 66 This weeks Editor: Cleve Moler Today's Topics: Parallel and Vector Algorithms Ulrich Hornung Address Change GAUSS and PC-MATLAB B. and K. Datta visit UCSD Chicago Conference Honoring Jim Douglas Domain Decomposition Conference, UCLA, Jan 14-16, 1988. Availability of SCHEDULE package on Netlib ---------------------------------------------------------------------- Date: 13 Aug 1987 5:03 PM CST From: FLORIAN A. POTRA <BLAFAPWY%UIAMVS.BITNET@wiscvm.wisc.edu> To: NA@SCORE.STANFORD.EDU Subject: Parallel and Vector Algorithms This Fall I will be teaching a graduate course on "Parallel & Vector Algorithms in Scientific Computing". During the semester the students will have to complete individual projects on the Alliant FX/8 and the Encore Multimax machines at the High Speed Computing Facility of the University of Iowa. I would like to introduce my students to the newest algorithms which are suitable for these machines. Therefore I would appreciate very much any recent reprints and/or preprints of papers related to the subject of the course as well as any computer programmes in FORTRAN or C. My address is: Florian A. Potra e-mail: Dept. of Mathematics na.potra@score.stanford.edu University of Iowa Iowa City, IA 52242 tel: (319)335-0776 With many thanks in advance, F.A.P. P.S. We are not able to print TEX documents at U of I. ------------------------------ DATE: 14 AUG 87 17:03 CET TO: NA@SCORE.STANFORD.EDU FROM: K1401AE%DM0LRZ01.BITNET@wiscvm.wisc.edu SUBJECT: Ulrich Hornung Address Change ULRICH HORNUNG SCHI P.O. BOX 1222 D-8014 NEUBIBERG WEST GERMANY TELEPHONE: 089-6004-3386 TELEX: 5215800 BW D (MUNICH) ARPA: NA.HORNUNG AT SCORE.STANFORD.EDU SUBJ.: TEMPORARY CHANGE OF ADDRESS DEAR COLLEAGUE, PLEASE NOTE THAT MY ADDRESS FROM SEPT. 1, 1987 THROUGH MAY 15, 1988 WILL BE DEPARTMENT OF MATHEMATICS ARIZONA STATE UNIVERSITY TEMPE, AZ 85287 U.S.A. TELEFONE 001-602-965-3951 TELEX 1561058 ASU UT ARPA: NA.HORNUNG AT SCORE.STANFORD.EDU SINCERELY ULRICH HORNUNG ------------------------------ Date: Wed, 19 Aug 87 10:49:21 BST From: Nick Higham (Manchester) <na.higham@score.stanford.edu> To: na@score.stanford.edu Subject: GAUSS and PC-MATLAB Matrix Computations on a PC: GAUSS and PC-MATLAB. In response to the recent request for more contributions to NA Digest I offer some thoughts on two interactive matrix computation packages for MSDOS machines (i.e. PCs): GAUSS and PC-MATLAB. This is not a comprehensive review, but rather brief notes aimed at informing people who work in matrix computations about some of the relative merits of these two packages. All opinions are entirely my own. Both packages require the presence of an 80(2)87 numeric co-processor. The 80(2)87 implements IEEE standard floating point arithmetic. It provides two precisions: 32 bit single precision (about 7 decimal digits), and 64 bit double precision (about 16 digits). All its internal computations are carried out in 80 bit extended precision. MATLAB should be familiar to readers of NA digest (see, e.g., the first part of Axel Ruhe's recent note). As regards PCs it exists in a public domain version, the original MATLAB (written in Fortran), and in a more recent, much faster and more extensive version, PC-MATLAB (written in C), which is marketed by The MathWorks. MATLAB is also available on SUNs and VAXes, in a version called PRO-MATLAB. GAUSS, from Aptech systems, is broadly similar to PC-MATLAB in its aims and features, but it is more strongly oriented towards the statistical analysis of data, and the manipulation of large sets of data stored on disk. Judging from the brochures it seems to be popular with statisticians, social scientists and economists. It exists only in a PC version and was first released in 1984/5. Both packages are advertised in the July 1987 issue of SIAM News and I refer you there for further information on prices, etc. Features in common to both packages: Use double precision. Seem to run the PC & co-processor at near maximum efficiency (both solve a 90x90 system Ax=b in under 30 seconds on a PC-AT). Interactive, and procedure driven modes. Include main linear system and eigenvalue routines from LINPACK and EISPACK. Powerful matrix syntax. Any single matrix is limited to 8192 elements (thus n<=90 for an nxn matrix) since it must fit into one 64K MSDOS segment. Can execute DOS commands from within the program. Contrasts: PC-MATLAB GAUSS --------- ----- Intrinsic complex arithmetic. Real arithmetic (some facilities for complex by working with real and imag. parts separately). Full help screens and demo programs. Neither - rather unfriendly to the new user. Line editor based, with capability Full screen editor - roam around the to recall and edit previous line. screen at will. Execute a sequence of commands contained within any marked area of the screen. Recall previous screen. Save screen to disk. Main program resides in memory. Uses disk overlays. FOR loops. No FOR loops - must be simulated using a DO loop. Vector elements referenced as X[i]. X[i,1] - must use two subscripts. Access to Hessenberg and Schur These absent - could be added by linking decompositions, matrix exponential in compiled Fortran code, but this not and square root, and QZ algorithm. for the feint-hearted! Square systems solved using Uses specially-written Crout/Cholesky LINPACK's DGECO/DPOCO. routines which accumulate inner products Rectangular Ax=b solved using the in extended precision. Crout has two QR decomposition (DQRDC). versions: no pivoting & partial pivoting. Rectangular Ax=b solved by forming and solving the normal equations entirely in extended precision ("better" than QR if A is well-conditioned!). Fast screen handling. Slower screen handling, due to forced use of ANSI.SYS screen driver. Output format 5 or 15 digits, Very versatile Fortran-like control scaled fixed or floating point. over output format. Control System Toolbox adds new Various add-on modules available, written commands for control applications. in GAUSS programming language. E.g. for hi-res graphics (since only three primitive routines are built in), unconstrained nonlinear optimisation (quasi-Newton), nonlinear least squares, further stats and data handling. Extremely easy to use, comprehensive On my copy hi-res graphics require a and impressive graphics. The one CGA adaptor, so I haven't tried them. program supports all the main I believe other graphics cards are graphics adaptors. supported in the latest version. Both packages include powerful programming languages (compiling to some forms of intermediate code) with procedures and functions. Both enable the user to to extend or customise the language using procedures. PC-MATLAB's syntax is the more elegant and concise, but GAUSS has more built-in commands, especially regarding statistical functions and handling "datasets" on disk. I find that the GAUSS compiler is fussy about syntax and its error messages are often extremely unhelpful. GAUSS offers great flexibility over configuration (how memory is partitioned between symbols and variables, which functions are pre- loaded etc.) but at the price of complexity. Summary: Both packages are extremely useful aids to research in matrix computations, for "n <=90". I recommend both: each has some advantages over the other. PC-MATLAB is the easier to learn and the more user-friendly, and would probably be the first choice of most numerical analysts; GAUSS can be quicker and more convenient for the "power user". (Can anyone offer any comparative comments on PC Fortrans?) Nick Higham na.higham@score.stanford.edu Department of Mathematics University Of Manchester Manchester M13 9PL England ------------------------------ Date: Wed, 19 Aug 87 16:47:56 cdt From: niuvax!dattab@anl-mcs.ARPA Apparently-To: na@SCORE.STANFORD.EDU Subject: B. and K. Datta visit UCSD We will be spending our sabbatical year at University of California San Diego (Mathematics Department) starting September 1, 1987 until June 1988. We will continue to receive our NA e-mail there. Biswa and Karabi Datta ------------------------------ Date: Wed, 19 Aug 87 17:57:03 CDT From: "Douglas N. Arnold" <arnold%s1.ima.umn.edu@umn-cs.cs.umn.edu> To: NA@score.stanford.edu Subject: Chicago Conference Honoring Jim Douglas ***** Final Announcement and Program ***** For more information write or phone the Conference Secretary, Department of Mathematics, University of Chicago, 5734 University Avenue, Chicago, Illinois 60637; phone: (312) 702-7100. ----------------------------------------------------------------- ADVANCES IN COMPUTATIONAL MODELLING AND NUMERICAL ANALYSIS ADVANCES IN COMPUTATIONAL MODELLING AND NUMERICAL ANALYSIS --- A conference in honor of Jim Douglas, Jr. on the occasion of his 60th birthday --- September 10, 11, and 12, 1987 --- Kent Hall 1020-24 East 58 Street Chicago, Illinois 60637 (312) 702-7100 +-----------------------------------------------+ | Thursday, September 10, Kent Hall 120 | +-----------------------------------------------+ Morning session chairman: Richard Ewing 9:30 A.M. Ivo Babuska Finite element methods for elliptic University of Maryland problems with piecewise analytic data 11:00 A.M. Vidar Thomee Numerical solution of parabolic Chalmers Institute of Tech. integro-differential equations Afternoon session chairman: Richard Falk 2:00 P.M. James Bramble Some remarks on mixed methods Cornell University 4:00 P.M. Jean-Claude Nedelec Homogenization of the problem of eddy Ecole Polytechnique, Paris currents in a transformer core 6:30 P.M. Wine and informal buffet at Douglas residence, 5531 S. Kimbark Avenue +-----------------------------------------------+ | Friday, September 10, Kent Hall 107 | +-----------------------------------------------+ Morning session chairman: John Osborn 9:00 A.M. Joachim Nitsche Finite element methods for conformal Albert-Ludwigs-U., Freiburg mappings 10:30 A.M. Mitchell Luskin Numerical results for liquid crystals University of Minnesota Afternoon session chairman: Donatella Marini 1:00 P.M. Roland Glowinski Numerical methods for the Hamilton- University of Houston Jacobi equations 2:30 P.M. Mary Wheeler Modelling of subsurface transport Rice University 4:00 P.M. Douglas Arnold Recent resultson the Mindlin-Reissner University of Maryland plate equations 6:00 P.M. Banquet in honor of Jim Douglas at the Quadrangle Club, 1155 E 57 St. Cash bar from 6:00-7:00, dinner at 7:00 +-----------------------------------------------+ | Saturday, September 10, Kent Hall 120 | +-----------------------------------------------+ Morning session chairman: Raymond Johnson 9:00 A.M. Herb Keller To be announced CalTech 10:30 A.M. Craig Douglas Recent advances in parallel multigrid I.B.M. Watson Research Center Afternoon session chairman: Paul Saylor 1:00 P.M. Franco Brezzi To be announced Universita di Pavia, Italy 2:30 P.M. Jacques-Louis Lions Exact controllability of distributed College de France, Paris systems. New results. Open problems. ------------------------------ Date: Wed, 19 Aug 87 18:24:07 PDT From: Tony Chan <chan@MATH.UCLA.EDU> To: na.dis@score.stanford.edu Subject: Domain Decomposition Conference, UCLA, Jan 14-16, 1988. SECOND ANNOUNCEMENT AND CALL FOR PAPERS 2nd INTERNATIONAL SYMPOSIUM ON DOMAIN DECOMPOSITION METHODS University of California, Los Angeles, January 14 - 16, 1988. THEME Domain Decomposition is a class of methods for solving mathematical physics problems by decomposing the physical domain into smaller subdomains and obtaining the solution by solving smaller problems on these subdomains. The motivation may be : the ability to use different mathematical models and approximation methods in different subdomains, use of fast direct methods in subdomains, memory limitations of the computer and suitability for implementation on parallel computers. Applications can be found in many areas of scientific computing and related industrial applications, such as computational fluid dynamics and structural mechanics. This symposium is a sequel to the First International Symposium on Domain Decomposition Methods held in Paris in January, 1987. The aim is to bring together the leading researchers in this rapidly expanding and highly interdisciplinary field to survey and review the progress that has been made since the last symposium. There will be approximately 25 invited papers and a limited contributed papers/poster session. In selecting invited and contributed papers, the organizing committee will try to keep a balance between the mathematical development, associated algorithms and their implementation on parallel computers, and industrial applications. ORGANIZING COMMITTEE James Bramble (Cornell), Tony F. Chan (UCLA), Roland Glowinski (Houston/INRIA), Jacques Periaux (GAMNI/SMAI/AMDBA), Olof Widlund (NYU). INFORMATION FOR AUTHORS Authors who wish to submit a paper should send 3 copies of an extended abstract (no longer than 2 pages and include preprints/reports if available) to : Prof. Tony F. Chan, Department of Mathematics, UCLA, 405 Hilgard Ave., Los Angeles, CA 90024. (Electronic mail: chan@math.ucla.edu). The deadline for submission is November 1, 1987. Acceptance will be notified before December 1, 1987. It is expected that selected papers will be published in a proceedings about June 1988. Authors should plan to have a full length paper ready by mid-March 1988. REGISTRATION Registration forms, travel and hotel information can be obtained by writing to: Elain Barth, Dept. of Mathematics, UCLA, 405 Hilgard Av., Los Angeles, CA 90024. Registration fee is US$100, which includes refreshment, abstracts and one banquet ticket. Student registration fee is US$50. Extra banquet ticket is US$25. RELATED CONFERENCES The annual AIAA meeting will be held in Reno, Nevada from January 11-14, 1988. The Third conference on Hypercube Concurrent Computers and Applications will be held the following week (January 19-20, 1988) in nearby Pasadena. SPONSOR Office of Naval Research under an University Research Initiative Grant. ------------------------------ Date: Thu, 20 Aug 87 12:24:43 cdt From: sorensen@anl-mcs.ARPA (Danny Sorensen) To: na.dis@score.stanford.edu Subject: Availability of SCHEDULE package on Netlib The SCHEDULE package for writing explicitly parallel programs in Fortran is now available on NETLIB. The purpose of this package is to provide a tool for constructing portable software for parallel computers. It allows one to express algorithms with synchronization requirements that are beyond the capabilities of loop based parallelism normally supported by vendors. It also provides the capability of dynamic spawning of processes at run time and interfaces well with existing Fortran libraries. The users code will port between the parallel computers that SCHEDULE is running on without change. In addition there is a graphics post processing capability that is useful for performance analysis and debugging. This trace facility runs on SUN workstations. The package is available on the following computers SUN VAX ALLIANT FX/8 SEQUENT BALANCE 21000 ENCORE MULTIMAX CRAY-2 FLEX (experimental) To get a copy of the code for one of the above machines send e-mail to netlib mail netlib@anl-mcs.arpa In the message body put send <machine name> from sched send trace from sched send uguide from sched where <machine name> is one of the following alliant balance cray2 encore flex sun trace is the graphics post processor for a SUN workstation and uguide is the current draft of the users guide. Dan Sorensen and Jack Dongarra sorensen@anl-mcs.arpa dongarra@anl-mcs.arpa Mathematics and Computer Science Division Argonne National Laboratory ------------------------------ End of NA Digest ************************** -------