findsame 0.1.0

Find duplicate files and directories using hashes and a Merkle tree

About

Find duplicate files and directories.

As other tools we use file hashes but additionally, we report duplicate directories as well, using a Merkle tree for directory hash calculation.

To increase performance, we use parallel hash calculation and optional limits on to-be-hashed data.

Install

From pypi:

$ pip3 install findsame

Dev install of this repo:

$ git clone ...
$ cd findsame
$ pip3 install -e .

The core part (package findsame and the CLI bin/findsame) have no external dependencies. If you want to run the benchmarks (see “Benchmarks” below), install some dependencies:

$ pip3 install -r requirements_benchmark.txt

Usage

usage: findsame [-h] [-b BLOCKSIZE] [-l LIMIT] [-L AUTO_LIMIT_MIN] [-p NPROCS]
                [-t NTHREADS] [-o OUTMODE] [-v]
                file/dir [file/dir ...]

Find same files and dirs based on file hashes.

positional arguments:
  file/dir              files and/or dirs to compare

optional arguments:
  -h, --help            show this help message and exit
  -b BLOCKSIZE, --blocksize BLOCKSIZE
                        blocksize in hash calculation, use units K,M,G as in
                        100M, 218K or just 1024 (bytes) [default: 256.0K]
  -l LIMIT, --limit LIMIT
                        read limit (bytes or 'auto'), if bytes then same units
                        as for BLOCKSIZE apply, calculate hash only over the
                        first LIMIT bytes, makes things go faster for may
                        large files, try 500K [default: None], use 'auto' to
                        try to determine the smallest value necessary
                        automatically
  -L AUTO_LIMIT_MIN, --auto-limit-min AUTO_LIMIT_MIN
                        start value for auto LIMIT calculation when --limit
                        auto is used [default: 8.0K]
  -p NPROCS, --nprocs NPROCS
                        number of parallel processes [default: 1]
  -t NTHREADS, --nthreads NTHREADS
                        threads per process [default: 4]
  -o OUTMODE, --outmode OUTMODE
                        1: json, 2: json with hashes [default: 1]
  -v, --verbose         enable verbose/debugging output

The output format is json, either with or without hashes (see --outmode). Use jq for pretty-printing. Example using the test suite data.

$ cd findsame/tests
$ findsame data | jq .
[
  {
    "dir:empty": [
      "data/dir2/empty_dir",
      "data/dir2/empty_dir_copy",
      "data/empty_dir",
      "data/empty_dir_copy"
    ],
    "file:empty": [
      "data/dir2/empty_dir/empty_file",
      "data/dir2/empty_dir_copy/empty_file",
      "data/empty_dir/empty_file",
      "data/empty_dir_copy/empty_file",
      "data/empty_file",
      "data/empty_file_copy"
    ]
  },
  {
    "dir": [
      "data/dir1",
      "data/dir1_copy"
    ]
  },
  {
    "file": [
      "data/file1",
      "data/file1_copy"
    ]
  },
  {
    "file": [
      "data/dir1/file2",
      "data/dir1/file2_copy",
      "data/dir1_copy/file2",
      "data/dir1_copy/file2_copy",
      "data/file2"
    ]
  },
  {
    "file": [
      "data/lena.png",
      "data/lena_copy.png"
    ]
  }
]

This is a json array (list) of objects (dictionaries) of same-hash files/dirs.

Note that currently, we skip symlinks.

Performance

Parallel hash calculation

By default, we use --nthreads equal to the number of cores. See “Benchmarks” below.

Limit data to be hashed

Static limit

Apart from parallelization, by far the most speed is gained by using --limit. Note that this may lead to false positives, if files are exactly equal in the first LIMIT bytes. Finding a good enough value can be done by trial and error. Try 500K. This is still quite fast and seems to cover most real-world data.

Automatic optimal limit

We have an experimental feature where we iteratively increase LIMIT to find the smallest possible value. In every iteration, we increase the last limit (see config.cfg.auto_limit_increase_fac) and with that re-calculate only the hash of files that have the same hash as others within the last LIMIT and check whether their new hashes are now different. This works but hasn’t been extensively benchmarked. The assumption is that a small number of iterations on a subset of all files (those reported equal so far) converges quickly and is still faster than a non-optimal LIMIT or even no limit at all when you have many big files (as in GiB).

Related options and defaults:

• --limit auto

• --auto-limit-min 8K = config.cfg.auto_limit_min

• config.cfg.auto_limit_increase_fac=2 (no cmd line so far)

Observations so far:

Convergence corner cases: When files are equal in a good chunk at file start and auto_limit_min is small, then the first few iterations show no change in files being equal (which we use to detect converged limit values). To circumvent early converge here, we iterate until the number of equal files changes. The worst case scenario is that auto_limit_min is already optimal. Since there is no way to determine that a priori, we will iterate until limit hits the biggest file size. That is why it is important to choose the start value small enough.

Start value: Don’t use very small start values such as 20 (that is 20 bytes), we found that this can converge to a local optimum (converged but too many equal files reported), depending in the structure of the headers of the files you compare. Stick with something like a small multiple of the blocksize of your file system (we use 8K).

Tests

Run nosetests3 (maybe apt install python3-nose before (Debian)).

Benchmarks

You may run the benchmark script to find the best blocksize and number threads and/or processes for hash calculations on your machine.

$ cd benchmark
$ ./clean.sh; ./benchmark.py
$ ./plot.py

This writes test files of various size to benchmark/files and runs a couple of benchmarks (runtime ~10 min for all benchmarks). Tune maxsize in benchmark.py to have faster tests or disable some benchmark functions.

Bottom line:

• blocksizes below 512 KiB (--blocksize 512K) work best for all file sizes on most systems, even though the variation to worst timings is at most factor 1.25 (e.g. 1 vs. 1.25 seconds)

• multithreading (-t/--nthreads): up to 2x speedup on dual-core box – very efficient, use NTHREADS = number of cores for good baseline performance (problem is mostly IO-bound)

• multiprocessing (-p/--nprocs): less efficient speedup, but on some systems NPROCS + NTHREADS is even a bit faster than NTHREADS alone, testing is mandatory

• we have a linear increase of runtime with filesize, of course

Tested systems:

• Lenovo E330, Samsung 840 Evo SSD, Core i3-3120M (2 cores, 2 threads / core)

• Lenovo X230, Samsung 840 Evo SSD, Core i5-3210M (2 cores, 2 threads / core)

  • best blocksizes = 256K

  • speedups: NPROCS=2: 1.5, NTHREADS=2..3: 1.9, no gain when using NPROCS+NTHREADS

• FreeNAS 11 (FreeBSD 11.0), ZFS mirror WD Red WD40EFRX, Intel Celeron J3160 (4 cores, 1 thread / core)

  • best blocksizes = 80K

  • speedups: NPROCS=3..4: 2.1..2.2, NTHREADS=4..6: 2.6..2.7, NPROCS=3..4,NTHREADS=4: 3

More usage examples

Output with hashes (-o 2, default is -o 1):

$ findsame data -o2 | jq . | head -n20
{
  "da39a3ee5e6b4b0d3255bfef95601890afd80709": {
    "dir:empty": [
      "data/dir2/empty_dir",
      "data/dir2/empty_dir_copy",
      "data/empty_dir",
      "data/empty_dir_copy"
    ],
    "file:empty": [
      "data/dir2/empty_dir/empty_file",
      "data/dir2/empty_dir_copy/empty_file",
      "data/empty_dir/empty_file",
      "data/empty_dir_copy/empty_file",
      "data/empty_file",
      "data/empty_file_copy"
    ]
  },
  "55341fe74a3497b53438f9b724b3e8cdaf728edc": {
    "dir": [
      "data/dir1",

In this case the output is one json object where all same-hash files/dirs are found at the same key (hash).

Note that the order of key-value entries in the output from both findsame and jq is random.

Post-processing is only limited by your ability to process json (using jq, Python, …).

Count the total number of all equals:

$ findsame data | jq '.[]|.[]|.[]' | wc -l

A common task is to find only groups of equal dirs:

$ findsame data | jq '.[]|select(.dir)|.dir'
[
  "data/dir1",
  "data/dir1_copy"
]

This and all other jq commands work for both outmodes (-o 1, -o 2). Now only the files:

$ findsame data | jq '.[]|select(.file)|.file'
[
  "data/dir1/file2",
  "data/dir1/file2_copy",
  "data/dir1_copy/file2",
  "data/dir1_copy/file2_copy",
  "data/file2"
]
[
  "data/lena.png",
  "data/lena_copy.png"
]
[
  "data/file1",
  "data/file1_copy"
]

Another task is to find the first or all but the first elements in a group of same-hash files/dirs.

Find first element:

$ findsame data | jq '.[]|.[]|[.[0]]'
[
  "data/lena.png"
]
[
  "data/dir2/empty_dir"
]
[
  "data/dir2/empty_dir/empty_file"
]
[
  "data/dir1/file2"
]
[
  "data/file1"
]
[
  "data/dir1"
]

or w/o the length-1 list:

$ findsame data | jq '.[]|.[]|.[0]'
"data/dir2/empty_dir"
"data/dir2/empty_dir/empty_file"
"data/dir1/file2"
"data/lena.png"
"data/file1"
"data/dir1"

All but first:

$ findsame data | jq '.[]|.[]|.[1:]'
[
  "data/dir1_copy"
]
[
  "data/lena_copy.png"
]
[
  "data/dir1/file2_copy",
  "data/dir1_copy/file2",
  "data/dir1_copy/file2_copy",
  "data/file2"
]
[
  "data/dir2/empty_dir_copy/empty_file",
  "data/empty_dir/empty_file",
  "data/empty_dir_copy/empty_file",
  "data/empty_file",
  "data/empty_file_copy"
]
[
  "data/dir2/empty_dir_copy",
  "data/empty_dir",
  "data/empty_dir_copy"
]
[
  "data/file1_copy"
]

And w/o lists:

$ findsame data | jq '.[]|.[]|.[1:]|.[]'
"data/file1_copy"
"data/dir1/file2_copy"
"data/dir1_copy/file2"
"data/dir1_copy/file2_copy"
"data/file2"
"data/lena_copy.png"
"data/dir2/empty_dir_copy/empty_file"
"data/empty_dir/empty_file"
"data/empty_dir_copy/empty_file"
"data/empty_file"
"data/empty_file_copy"
"data/dir2/empty_dir_copy"
"data/empty_dir"
"data/empty_dir_copy"
"data/dir1_copy"

The last one can be used, for example, to delete all but the first in a group of equal files/dirs, e.g.:

$ findsame data | jq '.[]|.[]|.[1:]|.[]' | xargs cp -rvt duplicates/

Other tools

• fdupes

• findup from fslint