The only configuration parameter that must be provided is the hash table size. Other parameters are optional or hardcoded, and the defaults are reasonable in most cases.
Hash table entries are 16 bytes per data block. The hash table stores the most recently read unique hashes. Once the hash table is full, each new entry added to the table evicts an old entry. This makes the hash table a sliding window over the most recently scanned data from the filesystem.
Here are some numbers to estimate appropriate hash table sizes:
unique data size | hash table size |average dedupe extent size
1TB | 4GB | 4K
1TB | 1GB | 16K
1TB | 256MB | 64K
1TB | 128MB | 128K <- recommended
1TB | 16MB | 1024K
64TB | 1GB | 1024K
Notes:
-
If the hash table is too large, no extra dedupe efficiency is obtained, and the extra space wastes RAM. If the hash table contains more block records than there are blocks in the filesystem, the extra space can slow bees down. A table that is too large prevents obsolete data from being evicted, so bees wastes time looking for matching data that is no longer present on the filesystem.
-
If the hash table is too small, bees extrapolates from matching blocks to find matching adjacent blocks in the filesystem that have been evicted from the hash table. In other words, bees only needs to find one block in common between two extents in order to be able to dedupe the entire extents. This provides significantly more dedupe hit rate per hash table byte than other dedupe tools.
-
There is a fairly wide range of usable hash sizes, and performances degrades according to a smooth probabilistic curve in both directions. Double or half the optimium size usually works just as well.
-
When counting unique data in compressed data blocks to estimate optimum hash table size, count the uncompressed size of the data.
-
Another way to approach the hash table size is to simply decide how much RAM can be spared without too much discomfort, give bees that amount of RAM, and accept whatever dedupe hit rate occurs as a result. bees will do the best job it can with the RAM it is given.
It is difficult to predict the net effect of data layout and access patterns on dedupe effectiveness without performing deep inspection of both the filesystem data and its structure--a task that is as expensive as performing the deduplication.
-
Compression on the filesystem reduces the average extent length compared to uncompressed filesystems. The maximum compressed extent length on btrfs is 128KB, while the maximum uncompressed extent length is 128MB. Longer extents decrease the optimum hash table size while shorter extents increase the optimum hash table size because the probability of a hash table entry being present (i.e. unevicted) in each extent is proportional to the extent length.
As a rule of thumb, the optimal hash table size for a compressed filesystem is 2-4x larger than the optimal hash table size for the same data on an uncompressed filesystem. Dedupe efficiency falls dramatically with hash tables smaller than 128MB/TB as the average dedupe extent size is larger than the largest possible compressed extent size (128KB).
-
Short writes or fragmentation also shorten the average extent length and increase optimum hash table size. If a database writes to files randomly using 4K page writes, all of these extents will be 4K in length, and the hash table size must be increased to retain each one (or the user must accept a lower dedupe hit rate).
Defragmenting files that have had many short writes increases the extent length and therefore reduces the optimum hash table size.
-
Time between duplicate writes also affects the optimum hash table size. bees reads data blocks in logical order during its first pass, and after that new data blocks are read incrementally a few seconds or minutes after they are written. bees finds more matching blocks if there is a smaller amount of data between the matching reads, i.e. there are fewer blocks evicted from the hash table. If most identical writes to the filesystem occur near the same time, the optimum hash table size is smaller. If most identical writes occur over longer intervals of time, the optimum hash table size must be larger to avoid evicting hashes from the table before matches are found.
For example, a build server normally writes out very similar source code files over and over, so it will need a smaller hash table than a backup server which has to refer to the oldest data on the filesystem every time a new client machine's data is added to the server.
The --scan-mode option affects how bees schedules worker threads
between subvolumes. Scan modes are an experimental feature and will
likely be deprecated in favor of a better solution.
Scan mode can be changed at any time by restarting bees with a different mode option. Scan state tracking is the same for all of the currently implemented modes. The difference between the modes is the order in which subvols are selected.
If a filesystem has only one subvolume with data in it, then the
--scan-mode option has no effect. In this case, there is only one
subvolume to scan, so worker threads will all scan that one.
Within a subvol, there is a single optimal scan order: files are scanned in ascending numerical inode order. Each worker will scan a different inode to avoid having the threads contend with each other for locks. File data is read sequentially and in order, but old blocks from earlier scans are skipped.
Between subvols, there are several scheduling algorithms with different trade-offs:
Scan mode 0, "lockstep", scans the same inode number in each subvol at close to the same time. This is useful if the subvols are snapshots with a common ancestor, since the same inode number in each subvol will have similar or identical contents. This maximizes the likelihood that all of the references to a snapshot of a file are scanned at close to the same time, improving dedupe hit rate and possibly taking advantage of VFS caching in the Linux kernel. If the subvols are unrelated (i.e. not snapshots of a single subvol) then this mode does not provide significant benefit over random selection. This mode uses smaller amounts of temporary space for shorter periods of time when most subvols are snapshots. When a new snapshot is created, this mode will stop scanning other subvols and scan the new snapshot until the same inode number is reached in each subvol, which will effectively stop dedupe temporarily as this data has already been scanned and deduped in the other snapshots.
Scan mode 1, "independent", scans the next inode with new data in each subvol. Each subvol's scanner shares inodes uniformly with all other subvol scanners until the subvol has no new inodes left. This mode makes continuous forward progress across the filesystem and provides average performance across a variety of workloads, but is slow to respond to new data, and may spend a lot of time deduping short-lived subvols that will soon be deleted when it is preferable to dedupe long-lived subvols that will be the origin of future snapshots. When a new snapshot is created, previous subvol scans continue as before, but the time is now divided among one more subvol.
Scan mode 2, "sequential", scans one subvol at a time, in numerical subvol
ID order, processing each subvol completely before proceeding to the
next subvol. This avoids spending time scanning short-lived snapshots
that will be deleted before they can be fully deduped (e.g. those used
for btrfs send). Scanning is concentrated on older subvols that are
more likely to be origin subvols for future snapshots, eliminating the
need to dedupe future snapshots separately. This mode uses the largest
amount of temporary space for the longest time, and typically requires
a larger hash table to maintain dedupe hit rate.
Scan mode 3, "recent", scans the subvols with the highest min_transid
value first (i.e. the ones that were most recently completely scanned),
then falls back to "independent" mode to break ties. This interrupts
long scans of old subvols to give a rapid dedupe response to new data,
then returns to the old subvols after the new data is scanned. It is
useful for large filesystems with multiple active subvols and rotating
snapshots, where the first-pass scan can take months, but new duplicate
data appears every day.
The default scan mode is 1, "independent".
If you are using bees for the first time on a filesystem with many existing snapshots, you should read about snapshot gotchas.
By default, bees creates one worker thread for each CPU detected.
These threads then perform scanning and dedupe operations. The number of
worker threads can be set with the --thread-count and --thread-factor
options.
If desired, bees can automatically increase or decrease the number
of worker threads in response to system load. This reduces impact on
the rest of the system by pausing bees when other CPU and IO intensive
loads are active on the system, and resumes bees when the other loads
are inactive. This is configured with the --loadavg-target and
--thread-min options.
bees can be made less chatty with the --verbose option.