home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky comp.arch:10639 comp.lang.misc:3600
- Newsgroups: comp.arch,comp.lang.misc
- Path: sparky!uunet!zaphod.mps.ohio-state.edu!darwin.sura.net!Sirius.dfn.de!Urmel.Informatik.RWTH-Aachen.DE!martin
- From: martin@math.rwth-aachen.de ( Martin Schoenert)
- Subject: Re: Hardware Support for Numeric Algorithms
- Message-ID: <martin.721555689@bert>
- Sender: news@Urmel.Informatik.RWTH-Aachen.DE (Newsfiles Owner)
- Nntp-Posting-Host: bert.math.rwth-aachen.de
- Organization: Rechnerbetrieb Informatik / RWTH Aachen
- References: <BwJ4uz.1rA@rice.edu> <1992Oct23.004313.29196@ntuix.ntu.ac.sg> <1992Oct29.153514.22927@yrloc.ipsa.reuter.COM> <1992Nov5.202412.7266@linus.mitre.org> <1992Nov10.153705.27804@yrloc.ipsa.reuter.COM>
- Date: 12 Nov 92 08:08:09 GMT
- Lines: 61
-
- rbe@yrloc.ipsa.reuter.COM (Robert Bernecky) writes:
-
- I write programs for correctness and maintainability first, THEN worry
- about little details such as performance. If you create something that's
-
- bs@gauss.mitre.org (Robert D. Silverman) replied:
-
- You've probably never done any large scale computing.
-
- I measure the run time of my algorithms in MIPS-YEARS. A MIPS-YEAR
- is a 1 MIPS machine running for 1 year, or approximately 3.1 x 10^13
- instructions. Jobs that take several hundred MIPS-YEARS are not uncommon.
-
- rbe@yrloc.ipsa.reuter.COM (Robert Bernecky) replied:
-
- Bad guess, unless Crays, 3090s (Yes, I know they barely count) and
- a bit of CM-2 don't count.
- Perhaps your code REALLY is MIPS-years in complexity. OR, it could
- be that failure to design properly leads to that complexity, whereas
- careful thought could reduce it to months or less. Since I don't
- understand your problem, I can only speculate. However, I DO know
- that algorithmic excellence is the way to get speed, NOT dinking
- around with what you perceive to be a bottleneck at the bit-diddling
- level. AFTER you have the algorithm right, THEN you start tweaking
- bits. Optimization before design is usually a bad idea.
-
- What you say about working on the algorithm first and only then
- optimizing the code is in principle right. But as a matter of fact it
- isn't a counterargument.
-
- Bob is talking about factorizations of large integers (at least I assume
- so). There are several extremly brilliant mathematicians doing nothing
- but trying to find an even better algorithm to factorize integers. And
- there have been several new algorithms in the last years (elliptic curve
- method, [multiple] polynomial quadratic sieve, number field sieve).
- Those algorithms have been implemented very carefully, with all kinds of
- optimizations. Still if you take such an implementation and apply it to
- the currently most wanted factorizations they *will* run for MIPS years
- (e.g. Lenstra, Lenstra, Manasse, and Pollard estimated that the
- factorization of F_9 = 2^512 + 1 took about 340 MIPS years).
-
- bs@gauss.mitre.org (Robert D. Silverman) continued:
-
- Needless to say, there are people who have to worry about speed FIRST,
- and that other considerations are (almost) irrelevent.
-
- rbe@yrloc.ipsa.reuter.COM (Robert Bernecky) replied:
-
- Are you saying that speed is more important than correct answers?
- {I don't think so, but I though I had better ask.}
-
- No, I don't think that this is what Bob is talking about. However in the
- area of integer factorization correctness is easy to achieve, because you
- can check your answers so easily. But reasonable efficiency can be
- *very* difficult to obtain.
-
- Martin.
-
- -- .- .-. - .. -. .-.. --- ...- . ... .- -. -. .. -.- .-
- Martin Sch"onert, Martin.Schoenert@Math.RWTH-Aachen.DE, +49 241 804551
- Lehrstuhl D f"ur Mathematik, Templergraben 64, RWTH, D 51 Aachen, Germany
-