Rodolfo Gonzalez Gonzalez put forth on 2/20/2010 12:18 AM:
I used to have the maildirs on ReiserFS and never had a problem with it, but given the current state of that FS and that I weren't really comfortable with it, I'll give XFS a try for the maildir array and the postfix queue partition. After formating, I got 4.6 Tb of usable space, which makes me happy, and also the dynamic inode allocation.
http://en.wikipedia.org/wiki/XFS
Like I said, it's a very mature high performance journaled FS with many enterprise level features, dynamic inode allocation being one of many. It was introduced by SGI in 1994 and has been in constant development since then. It was ported to Linux around 2000 and introduced into the mainline kernel in 2.4.
It is the only filesystem ever used on SMP servers from 128+ CPUs up to 1024 CPUs. This is because SGI is the only company to ever offer SMP systems beyond 128 CPUs. They are actually ccNUMA, not SMP, but the programming model is SMP, because every CPU in the machine can directly address memory in any NUMA node in the system. The only practical difference between ccNUMA and a true SMP is the memory latency.
Obviously, scalability and the ability to manipulate very large filesystems with large numbers of files is required for such massive machines. The Columbia supercomputer at the NASA Ames facility consists of 20 such machines, each with 512 CPUs. The system has a 1 Peta Byte (raw) RAID subsystem formatted with CXFS, the clustered version of XFS.
XFS scales very well. ;)
I've been fan of SGI for a long time. I could never afford/justify one of their machines. I'm so glad they open sourced XFS and are sharing this fantastic filesystem with the rest of us who could never afford their gear. Many would agree with me if I said it is hands down the overall best *nix filesystem available for most workloads. It's not suitable on Linux for /boot or /, but for just about everything else it is king of the hill.
-- Stan