On Mon, May 23, 2005 at 03:22:37AM -0700, Dan Hollis wrote: > On Mon, 23 May 2005, Arjan van de Ven wrote: > > On Mon, May 23, 2005 at 03:02:23AM -0700, Dan Hollis wrote: > > > On Mon, 23 May 2005, Arjan van de Ven wrote: > > > > > Both. > > > > > Its not just for large directories, reiserfs did much better with many > > > > > small files too (typical of news and mailservers). > > > > Hmm. that surprises me for htree enabled filesystems (note that if you > > > > create an FS with an old distro and then put 2.6 on it it doesn't use htree) > > > Remember reiserfs was designed from the bottom up to work quickly with > > > lots of _small files_. So it does what it was designed to do well. That > > > this happens to be the typical workload of news servers and mailservers > > > is a happy coincidence. > > yet a design goal leads to a technological implementation, and I'm wondering > > what is a causing factor for ext3 to not be roughly equally fast. ext3 with > > htree should in principle not be bad at lots of small files. at all. There > > is no inherent bias in ext3 towards bigger files (well except when you count > > not having tail merging; tail merging will give you gain in the case of a > > read-mostly lots-of-small-files case) > > I think some of this has to do with the fact that ext3 is fundamentally a > very, very, very old filesystem. Its some of the oldest code still in the > kernel iirc. By no means true. > Some of the assumptions made in the original design may no > longer be so relevant over a decade later. ext2 was in use when everyone > was still on PIO and C/H/S :-) actually that isn't the case. sure there was an ext2 way back. But it moved forward quite a bit (htree is an example of that). > reiserfs by comparison is brand spanking > new. reiserfs3 isn't so new though, it is like 6 years old. reiserfs4 is entirely unrelated to reiserfs3. The biggest difference between reiserfs3 and ext3 is that reiserfs3 is more tree based on disk and ext3 uses blockgroups on disk. For directories they nowadays both uses trees (since ext3 grew htrees). I haven't seen conclusive evidence either way which is architectually better; trees tend to be a bit more fragile but can do certain things better (but other things really poor, if the tree has to rebalance all the time, in addition trees tend to cause more fragementation longer term). Also there is a difference between the on disk storage of metadata and the in memory one. If you want to use reiserfs, by all means be my guest. Some people only think they have performance gains with it, others really do. Reiserfs in fedora tends to be "lightly tested" at best though, so make sure you test it hard yourself before going into production.