Import performance with David Kimdon's full-text patch vs vdelta

Just went and tested the performance of David Kimdon's patch to
eliminate vdelta overhead on imports.

http://subversion.tigris.org/servlets/ReadMsg?list=dev&msgNo=47715

I ran two tests, one to import the Mozilla source code tree and the
other to import the Linux source code tree. The results are pretty much
as expected, a pretty decent performance win. I am not sure the speedup
is significant enough to warrant using this patch over a fast
connection, but I would definitely recommend using this patch for large
ra_local operations.

The (x) comparison is against the base "SVN without patch" times.

I've omitted the Perforce and Arch times for the Mozilla import because
I could not do these operations in a single (recursive) call like SVN
and CVS (even though I could get these times via xargs).

Here is the the test setup:
SVN without patch - 93 seconds
SVN with patch - 71 seconds (29

Linux 2.6.0-X8664-1 i686
AMD Athlon(tm) 64 Processor 3000+ (2.0GHZ, 512KB cache)
DDR400 1GB
38GB Serial ATA Raptor @ 10,000 RPM (WDC WD360GD-00FNA0)
Ext3 with -o journal_data_writeback
Subversion trunk @ 8151
Repository created with --bdb-txn-nosync

Mozilla 1.6b source import

SVN without patch - 158 seconds
SVN with patch - 139 seconds (1.14x)
CVS - 10 seconds (15.80x)
Perforce - failed (exceeded shell parameter limit)
Arch - failed (exceeded shell parameter limit)

Linux 2.6.0 source import

SVN without patch - 93 seconds
SVN with patch - 71 seconds (1.31x)
CVS - 8 seconds (11.63x)
Perforce - 165 seconds (0.56x)
Arch - 67 seconds (x1.39)

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@subversion.tigris.org
For additional commands, e-mail: dev-help@subversion.tigris.org
Received on Mon Jan 5 06:15:09 2004