The patch is a cleanup. It saves 85 bytes on x86.
size hashes.o*
text data bss dec hex filename
1333 16 0 1349 545 hashes.o
1248 16 0 1264 4f0 hashes.o.new
I tested the patch on uniprocessor x86 on filesystems
created with:
yes "y" | mkreiserfs --format 3.6 -h tea /dev/hdc1
mount -t reiserfs -o defaults,noatime,notail /dev/hdc1 /fs1
The keyed_hash function has 20 less instructions on x86,
but performance improvement is very small.
The default hash is r5, so in the common case, the patch
doesn't do anything except save a few bytes.
Time to run "dbench-2.0 32" 4x and "dbench-2.0 64" 5x.
2.6.0-test1-ac2-r 8560 seconds (patched)
2.6.0-test1-ac2-t 8615 seconds
The patched version is about .5% faster.
Oddly, the 3rd dbench 32 run on the unpatched version had 2
dbench processes that stayed Sleeping. Killing the
dbench processes let everything continue. Because of that
one skewed run, I deleted the 3rd dbench time from the
"seconds" above for both kernels.
bonnie++-1.03a didn't have much different between
patched and unpatched except for sequential block i/o.
Sequential Output
------Block------
Kernel Size MB/sec %CPU Eff
reiserfs-2.6.0-test1-ac2-r 1024 19.22 75.0 25.63 (patched)
reiserfs-2.6.0-test1-ac2-t 1024 18.18 70.0 25.97
tiobench-0.3.3 improvement with the patch was also small.
--
Randy Hron
http://home.earthlink.net/~rwhron/kernel/bigbox.html
