Subj : Re: transactional memory question To : comp.arch,comp.programming.threads From : pg_nh Date : Tue Oct 04 2005 07:33 pm >>> On Tue, 04 Oct 2005 09:25:54 -0600, Anne & Lynn Wheeler >>> said: [ ... transactional memory, what about it? ... ] lynn> original 801 had a form ... it was used by journal lynn> filesystem for aix in rs/6000. [ ... ] a portable version lynn> of the journal filesystem code with explicit lock/logging lynn> changes in the code turned up the explicit lock/logging lynn> calls had lower overhead than transactional memory (at lynn> least in the journal filesystem case). Ah the usual interesting historical and technical detail from you. Thanks, as usual, and I am particularly interested in this as I just switched my Linux based PC to the latter day version of that =ABportable version of the journal filesystem code=BB, the JFS in recent Linux kernels. As an aside I switched because I was terrified by some informal file system tests I have done, when I discovered that performance degradation due to extensive use in 'ext3' can be quite apocalyptic: http://WWW.sabi.co.UK/Notes/anno05-3rd.html#050913 [ ... ] lynn> A Fault Tolerant Tightly Coupled Multiprocessor lynn> Architecture based on Stable Transactional Memory lynn> Authors: Michel Banatre and Philippe Joubert Banatre? There was a Banatre at UofRennes who wrote interesting stuff about Algol 68 long ago, including an utterly fascinating analysis of how much micro parallelism/locking (a topic that has come up recently in comp.arch IIRC) there was in his Algol 68 compiler taken as an example. Ahhh, I could not resist looking that up. That Banatre is Jean-Pierre Banatre, and perhaps Michel is a relation or a second generation geek. Which category is happening a lot (e.g. IIRC famously with Larry Wall, but I also met once this awesome Oz girl both whose parents are CompSci professors, and she got a CompSci degree and is doing a CompSci PhD too :->). However, a little web searching with the obvious keywords and the microparallelism/microlocking paper is: http://DOI.ACM.org/10.1145/359046.359055 "An event-driven compiling technique" where the technique used to resolve Algol 68's mysterious forward references (which IIRC normally require at least four left-to-right passes) is basically microthreads... As another note, I feel compelled to mention MC/CWU's Bohm's ''oscillating'' compiler passes (surely revealed to him by aliens), which cuts the four passes down to two, where the first pass compiler is left-to-right, and the second is right-to-left, which as the author notes in effect gives infinite right lookahead. I have never seen it mentioned anywhere else... Enough. :-) No, not enough -- I could not resist looking up about that Bohm, and he seems to be this guy: http://WWW.CS.ColoState.edu/~bohm/v.pdf and I did not know that he had also worked at V.U. Manchester legendary CompSci department, and on their legendary dataflow systems (I had the privilege many years ago to see a full dataflow machine actually working over there, a rare experience indeed). Sorry, but at least his paper list contains a lot of nice references on microparallelism (and perhaps locking too) so it is not totally offtopic. :-) [ ... ] .