snappy-start is a tool which takes a snapshot of a Linux program's process state after it has started up. It allows multiple instances of the program to be quickly launched from the snapshot.
This has two potential benefits:
-
Faster startup, if the program does a non-trivial amount of computation during startup.
-
Saving memory, because memory pages that the program writes to during startup will be shared between the instances.
First, build the tool by running make.sh
(which also runs some tests).
To create a snapshot:
./out/ptracer ./out/elf_loader PROG ARGS...
This creates a self-contained executable replay.out
in the current
directory. Run it to restore the snapshot.
The program will be snapshotted when it first calls an unhandled
syscall, such as getpid()
.
Example:
$ ./out/ptracer ./out/elf_loader /usr/bin/python tests/python_example.py
$ ./replay.out
Hello world, from restored Python process
The idea for this tool comes from Kenton Varda, who proposed using a
"record/replay" approach, using ptrace()
to monitor syscalls so that
they can later be replayed.
Mark Seaborn put together an initial working end-to-end implementation.
Kenton enhanced the implementation to produce a self-contained binary and extended it with support for more syscalls.