From kragen@dnaco.net Thu Sep 24 22:29:48 1998 Date: Thu, 24 Sep 1998 22:29:46 -0400 (EDT) From: Kragen X-Sender: kragen@pike To: "Robert G. Brown" cc: Shachar Tal , beowulf@cesdis1.gsfc.nasa.gov, extreme-linux@acl.lanl.gov Subject: Re: Cluster-wide overclocking... In-Reply-To: Message-ID: MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII X-Keywords: X-UID: 2140 Status: O X-Status: On Thu, 24 Sep 1998, Robert G. Brown wrote: > There are other cases where it would be a disaster. If you are trying > to certify the passwd file as being "unbreakable" for a client in a > million dollar banking system and some kid running crack turns around > and finds it on a NON overclocked system in ten seconds, Well, this is still a bad example, because no reputable security person will certify a passwd file as unbreakable. The bad guys always have better dictionaries than the good guys, because the bad guys packet-sniffed netcom for several months and they know what kinds of passwords people *really* choose. BTW, there was an article somewhere recently about an experimental machine in which all the components were defective. The machine was just massively redundant, and cheaper to build than a non-redundant machine of equivalent power. There was a pointer to it on slashdot.org in the last week or two, IIRC. > Really, one can actually make this completely mathematical given any > expected rate of failure and a reasonable knowledge of the classes of > failure that can occur, That's the kicker, though. Expected rates of failure in computers tend to be badly unpredictable; classes of failure tend to be difficult to enumerate. > Statistics is one where that is VERY DANGEROUS to make > assumptions based on naive frequency interpretations, Yeah. I thought about your Monte Carlo calculations; it seems to me that one or two errors could easily swamp the whole run, on many kinds of Monte Carlo calculations. If you're calculating an approximation of a value X lots of times and then averaging your approximations to get a better X, in particular. Suppose that X is really about a billion. One single-bit error in a floating-point exponent could result in estimating an X that's closer to 10^20 or so. Your original thought that occasional one-bit errors wouldn't matter seems like a naive assumption about frequency distributions. But maybe your Monte Carlo calculations are of a type that wouldn't be affected in quite this way. Kragen -- Kragen Sitaker The sages do not believe that making no mistakes is a blessing. They believe, rather, that the great virtue of man lies in his ability to correct his mistakes and continually make a new man of himself. -- Wang Yang-Ming