xcp
is a (partial) clone of the Unix cp
command. It is not intended as a
full replacement, but as a companion utility with some more user-friendly
feedback and some optimisations that make sense under certain tasks (see
below).
Warning: xcp
is currently beta-level software and almost certainly contains
bugs and unexpected or inconsistent behaviour. It probably shouldn't be used for
anything critical yet.
NOTE: xcp
requires Rust 1.36 or higher.
xcp
can be installed directly from crates.io
with:
cargo install xcp
xcp
is available on the Arch Linux User Repository. If you use an AUR helper, you can execute a command such as this:
yay -S xcp
- Displays a progress-bar, both for directory and single file copies. This can
be disabled with
--no-progress
. - On Linux it uses
copy_file_range
call to copy files. This is the most efficient method of file-copying under Linux; in particular it is filesystem-aware, and can massively speed-up copies on network mounts by performing the copy operations server-side. However, unlikecopy_file_range
sparse files are detected and handled appropriately. - Optional aggressive parallelism for systems with parallel IO. Quick experiments on a modern laptop suggest there may be benefits to parallel copies on NVMe disks. This is obviously highly system-dependent.
- Non-Linux Unix-like OSs (OS X, *BSD) are supported via fall-back operation (although sparse-files are not yet supported in this case).
- Optionally understands
.gitignore
files to limit the copied directories. - Optimised for 'modern' systems (i.e. multiple cores, copious RAM, and solid-state disks, especially ones connected into the main system bus, e.g. M.2).
- Optional native file-globbing.
- Conversion of files to sparse where appropriate, as with
cp
's--sparse=always
flag. - Aggressive sparseness detection with
lseek
.
- On Linux
copy_file_range()
requires a kernel version of 4.5 and onwards; if it is missingxcp
will fall-back to user-space copy. - On non-Linux OSs sparse-files are not supported (although could be added if supported by the OS).
- Assumes a 'modern' system with lots of RAM and fast, solid-state disks. In particular it is likely to thrash on spinning disks as it attempts to gather metadata and perform copies at the same time.
- Currently missing a lot of
cp
's features and flags, although these could be added.
Benchmarks are mostly meaningless, but to check we're not introducing too much
overhead for local copies, the following are results from a laptop with an NVMe
disk and in single-user mode. The target copy directory is a git checkout of the
Firefox codebase, having been recently gc'd (i.e. a single 4.1GB pack
file). fstrim -va
is run before each test run to minimise SSD allocation
performance interference.
- Single 4.1GB file copy, with the kernel cache dropped each run:
cp
: ~6.2sxcp
: ~4.2s
- Single 4.1GB file copy, warmed cache (3 runs each):
cp
: ~1.85sxcp
: ~1.7s
- Directory copy, kernel cache dropped each run:
cp
: ~48sxcp
: ~56s
- Directory copy, warmed cache (3 runs each):
cp
: ~6.9sxcp
: ~7.4s
xcp
uses copy_file_range
, which is filesystem aware. On NFSv4 this will result
in the copy occurring server-side rather than transferring across the network. For
large files this can be a significant win:
- Single 4.1GB file on NFSv4 mount
cp
: 378sxcp
: ~37s