Hi ! CL> Hi Jeff, CL> I tested your bitmap hinting patch over the weekend! >> >> Constantin - >> >> I've posted a few times to the list regarding a bitmap hinting patch that >> I've written up. My results show a pretty decent performance increase with >> respect to both fragmentation and general write performance stemming from >> the more efficient allocation of blocks. >> >> I'd be very interested to see how this patch performs on your benchmarks. >> It's very likely to be the default bitmap allocator in the coming weeks. >> >> The patch is available at >> ftp://ftp.suse.com/pub/people/jeffm/reiserfs/kernel-v2.4 >> >> Please note that it also requires my endian safe patches. These patches are >> stable, and have been included in the -ac tree for several months. They're >> currently waiting on Linus to include them in his tree. The endian patches >> are available in that same directory. >> >> Thanks. >>
CL> My tests indicate, that the overall performance in absolut numbers CL> is slightly smaller than the plain Reiser on a 2.4.10 kernel (call CL> it also equal when looking at statistics eg. mean values CL> and mean derivation). CL> Looking at relative values (eg. by deviding the measured values through CL> the first CL> measured value of the curve) it clearly is seen that the BMP improves CL> the behaviour of Reiser with increasing age. CL> For example the random read values drop to 60 % of the CL> orginal performance for the BMP and to 50 % for stock RFS but CL> absolut number are almost equal. CL> Watch the graph CL> http://www.informatik.uni-frankfurt.de/~loizides/reiserfs/gif/agetest/compare-rfs-bmp-absolut.png CL> for absolut values and CL> http://www.informatik.uni-frankfurt.de/~loizides/reiserfs/gif/agetest/compare-rfs-bmp.png CL> for relative (normed) values. Can you test reiserfs with no_unhashed_relocation mount option? It seems more useful for large partitions and when fs is >50% full. CL> Some more details are outlined here: CL> http://www.informatik.uni-frankfurt.de/~loizides/reiserfs/agetest.html#meas2 CL> Constantin -- Thanks, Alex.
