On Sun, Feb 04, 2007 at 05:30:18PM +0100, A Mennucc wrote: > Package: bsdiff > Version: 4.3-3 > Severity: normal > > hi > > I manage the debdelta project > http://tonelli.sns.it/pub/mennucc1/debdelta/ > > debdelta needs a 'delta compressor' as a backend; > and bsdiff would be one of the best choices, as shown > by my tests in > http://tonelli.sns.it/pub/mennucc1/debdelta/tests > .... but for a strange problem ... actually, two problems.... > > ---Problem number one: > sometimes bsdiff hangs forever (or, for more than > I am willing to wait). I am capable of extracting from debdelta the > exact "old" and "new" files that hang it; but unfortunately most of > those are not easily reproducible ... it is as if bsdiff is > triggering a bug in my CPU or in my memory. > > You find all my collected examples of "old" and "new" files in > http://tonelli.sns.it/pub/mennucc1/debdelta/slowest_bsdiff/ > > BTW: I run debdeltas in a amd64 Etch install; the CPU is a Athlon64 > 3000 . That host is quite stable, and never shows problems with any > other service. > > ---Problem number two: > on some short files, where my experience says that > 'bsdiff' should compute a diff in a matter of minutes, then 'bsdiff' > would slow down a lot. > > The strangest one is gcl_old gcl_new : once (~6 month ago) I ran > $ bsdiff gcl_old gcl_new /tmp/diff > and it took 8 hours to complete.... this is 10 times slower than > average. (I am currently running that command, it has gone on > for 5 hours). > > I currently prepend 'stopafter 600 TERM' to my 'bsdiff' invocation > ('stopafter' is a nice utility that was in Debian once; that > cmd stops bsdiff after 600 seconds, using SIGTERM). > This "fixes" my problem, but is not really a fix.
We should apply the following 4 patches from chromium - they are small, and, as part of Chromium, also presumably well tested: (1) https://chromium.googlesource.com/chromiumos/third_party/bsdiff/+/e2b85ce4ea246ca804ee2a9e18005c44e193d868 Replacing the custom suffix sorting implementation with -ldivsufsort. This brings down the time to generate a delta (for the proposed new format) for libreoffice-core 5.4~rc2-1->5.4-1 from 93 seconds down to 50 seconds. It also just involves replacing code with a library call, so that's nice as well. (2) https://chromium.googlesource.com/chromiumos/third_party/bsdiff/+/a055996c743add7a9558839276fd1e4994d16bd3 Speeds up a pathological case (3) https://chromium.googlesource.com/chromiumos/third_party/bsdiff/+/58146f74abd6b1b69693943195f37f4ac6a6acef Fixes a hang (4) https://chromium.googlesource.com/chromiumos/third_party/bsdiff/+/426e4aa1cbeb3c8a73002047d7a796ca8e5e17d4 Another pathological case where files differ by less than 8 bytes -- Debian Developer - deb.li/jak | jak-linux.org - free software dev | Ubuntu Core Developer | When replying, only quote what is necessary, and write each reply directly below the part(s) it pertains to ('inline'). Thank you.
