[
https://issues.apache.org/jira/browse/SVN-1966?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14920134#comment-14920134
]
Ivan Zhakov edited comment on SVN-1966 at 10/17/15 3:25 PM:
------------------------------------------------------------
By default GNU diff uses heuristics to shortcut the minimal diff algorithm,
when these kick in it means that the resulting diff is not know to be a minimal
one. libsvn_diff doesn't have such heuristics and always computes a minimal
diff; this looks likes one of those cases where it is very expensive to compute
a minimal diff. GNU diff's --minimal option disables the heuristics; use it on
the test files and the GNU diff time goes through the roof. This doesn't
really explain the memory use, although as far as I recall GNU diff makes use
of stack memory for some things that libsvn_diff does on the heap.
was (Author: philipm):
{noformat:nopanel=true}
By default GNU diff uses heuristics to shortcut the minimal diff algorithm, when
these kick in it means that the resulting diff is not know to be a minimal one.
libsvn_diff doesn't have such heuristics and always computes a minimal diff;
this looks likes one of those cases where it is very expensive to compute a
minimal diff. GNU diff's --minimal option disables the heuristics; use it on
the test files and the GNU diff time goes through the roof. This doesn't really
explain the memory use, although as far as I recall GNU diff makes use of stack
memory for some things that libsvn_diff does on the heap.
{noformat}
> libsvn_diff needs 'non-minimal-diff' mode.
> ------------------------------------------
>
> Key: SVN-1966
> URL: https://issues.apache.org/jira/browse/SVN-1966
> Project: Subversion
> Issue Type: Improvement
> Components: libsvn_diff
> Affects Versions: all
> Reporter: Ben Collins-Sussman
> Assignee: Sander Striker
> Priority: Critical
> Fix For: unscheduled
>
> Attachments: 1_xmlfiles.tar.gz
>
>
> I've got two OpenOffice XML files here, each about 9 megs in size. The
> second file was produced by simply adding a "2" to the beginning of every
> third line (...we used a tiny python script to do the transform.)
> When I compare the two files using GNU diff, it takes 14 seconds.
> When I compare the two files using libsvn_diff, it never finishes... or
> rather, I killed the process after 20 minutes when I noticed the memory
> footprint had grown to 128megs and was still growing. The footprint was only
> ~20 megs for the
> first 10 minutes, which is expected behavior (about 2x the size of the file),
> but I have no idea why the footprint started growing after that.
> I'm using the subversion/tests/libsvn_diff/diff-test binary, by the way.
> I'm attaching the two XML files to this mail for people to reproduce.
> Sander, any idea what's going on? I know that libsvn_diff is sometimes a
> *bit* slower than gdiff, but this scenario seems way out of whack. Cmpilato
> and I discovered this bug when a Collabnet customer attempted to run 'svn
> merge': the
> client started running svn_diff_diff3() on two 8 meg XML files similar to
> this one (with a similar density of changed lines)... and after 20 minutes,
> the server just timed-out and closed the socket... presumably because the
> socket filled up with the REPORT response and the client stopped reading from
> it.
> Anyway, I hope this is a simple bug in libsvn_diff...
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
