As my primary OS is not Plan9 but FreeBSD, I am interested in venti for archiving data on UFS filesystems.
For most purposes, vac does a good job, but it cannot preserve the full semantics of a Unix filesystem, especially not symbolic links. Also, vac does not work on a filesystem, but on a live file tree, where mounted other filesystems may yield unwanted results.
There are several ways to handle this drawback, one could put the data into a tar- or cpio-archive and vac this file. Other variants are copying the data into a temporary UFS filesystem and use vbackup from planports. Even the classic
Unix dump may be a solution.
To evaluate the deduplication performance, I did a little experiment. On an empty arenapartion, I started with a vac of a directory (in this case my clone of Noam's neoventi). The .git subdirectory contains a lot of uncompressible files of different sizes.
total arenas=5 active=1
total space=20,094,976 used=914,716
clumps=512 compressed clumps=292 data="" compressed data="" />
After that, I tried my candidates, and after each run I restarted with a refreshed arena partition.
1. dump (of course with blocksize 8k)
total space=20,094,976 used=1,749,986
clumps=694 compressed clumps=413 data="" compressed data="" />
2. vbackup with UFS 8k /1k fragments
total space=20,094,976 used=1,198,550
clumps=645 compressed clumps=412 data="" compressed data="" />
3. vbackup with UFS 8k and no fragments
total space=20,094,976 used=960,511
clumps=659 compressed clumps=436 data="" compressed data="" />
4. tar also did not compare well, because it operates in 512 byte blocks, so the chance of proper alignment with 8k lumps on venti is about 1/16.
It may be noted as interesting that I needed 4m for the non-fragmented ufs filesystem, but only 2m for the fragmented version, but the 1st fares much better in used space.
Some more info on this on http://pkeus.de/~wb/bench.out
