sendfile(2) is a system call which provides a "zero-copy" way of copying data from one file descriptor to another (a socket). The phrase "zero-copy" refers to the fact that all of the copying of data between the two descriptors is done entirely by the kernel, with no copying of data into userspace buffers. This is particularly useful when sending a file over a socket (e.g. FTP). The normal way of sending a file over a socket involves reading data from the file into a userspace buffer, then write that buffer to the socket via send() or sendall():
# how a file is typically sent import socket file = open("somefile", "rb") sock = socket.socket() sock.connect(("127.0.0.1", 8021)) while True: chunk = file.read(65536) if not chunk: break # EOF sock.sendall(chunk)
This copying of the data twice (once into the userland buffer, and once out from that userland buffer) imposes some performance and resource penalties. sendfile(2) syscall avoids these penalties by avoiding any use of userland buffers; it also results in a single system call (and thus only one context switch), rather than the series of read(2) / write(2) system calls (each system call requiring a context switch) used internally for the data copying.
import socket import os from sendfile import sendfile file = open("somefile", "rb") blocksize = os.path.getsize("somefile") sock = socket.socket() sock.connect(("127.0.0.1", 8021)) offset = 0 while True: sent = sendfile(sock.fileno(), file.fileno(), offset, blocksize) offset += sent if sent == 0: break # EOF
A simple benchmark
This benchmark script implements the two examples above and compares plain socket.send() and sendfile() performances in terms of CPU time spent and bytes transmitted per second resulting in sendfile() being about 2.5x faster. These are the results I get on my Linux 2.6.38 box, AMD dual-core 1.6 GHz:
|CPU time||28.84 usec/pass|
|transfer rate||359.38 MB/sec|
|CPU time||11.28 usec/pass|
|transfer rate||860.88 MB/sec|
When do you want to use it?
sendfile module provides a single function: sendfile().
sendfile.sendfile(out, in, offset, nbytes, header="", trailer="", flags=0)
Copy nbytes bytes from file descriptor in (a regular file) to file descriptor out (a socket) starting at offset. Return the number of bytes just being sent. When the end of file is reached return 0. On Linux, if offset is given as None, the bytes are read from the current position of in and the position of in is updated. headers and trailers are strings that are written before and after the data from in is written. In cross platform applications their usage is discouraged (send() or sendall() can be used instead). On Solaris, out may be the file descriptor of a regular file or the file descriptor of a socket. On all other platforms, out must be the file descriptor of an open socket. flags argument is only supported on FreeBSD.
Parameters for the flags argument, if the implementation supports it. They are available on FreeBSD platforms. See FreeBSD's man sendfile(2).
Differences with send()
- sendfile(2) works with regular (mmap-like) files only (e.g. you can't use it with a StringIO object).
- Also, it must be clear that the file can only be sent "as is" (e.g. you can't modify the content while transmitting). There might be problems with non regular filesystems such as NFS, SMBFS/Samba and CIFS. For this please refer to proftpd documentation.
- since the file is sent "as is" sendfile(2) can only be used with clear-text sockets (meaning SSL is not supported).
- OSError is raised instead of socket.error. The accompaining error codes have the same meaning though: EAGAIN, EWOULDBLOCK, EBUSY meaning you are supposed to retry, ECONNRESET, ENOTCONN, ESHUTDOWN, ECONNABORTED in case of disconnection. Some examples: benchmark script, test suite, pyftpdlib wrapper.
- sendfile(2) can be used with non-blocking sockets, meaning if you try to send a chunk of data over a socket fd which is not "ready" you'll immediately get EAGAIN (then you can retry later by using select(), epoll() or whatever).
- the regular file fd, on the other hand, can block.
This module works with Python versions from 2.5 to 3.X and later on it was integrated into Python 3 as os.sendfile() and socket.socket().sendfile() (see bpo-10882 and bpo-17552). The supported platforms are:
- Mac OSX
- Dragon Fly BSD
- Sun OS
- AIX (not properly tested)
pysendfile was originally written by Ben Woolley including Linux, FreeBSD and DragonFly BSD support. Later on Niklas Edmundsson took over maintenance and added AIX support. After a couple of years of project stagnation Giampaolo Rodola' took over maintenance and rewrote it from scratch adding support for:
- Python 3
- non-blocking sockets
- large file support
- Mac OSX
- Sun OS
- FreeBSD flag argument
- multiple threads (release GIL)
- a simple benchmark suite
- unit tests