FreeSec - NetBSD libcrypt replacement
David Burren <davidb@werj.com.au>
Release 1.0, March 1994
Document ref: $Id$
Description
===========
This library is a drop-in replacement for the libcrypt used in U.S. copies
of NetBSD, duplicating that library's functionality. A suite of verification
and benchmark tools is provided.
FreeSec 1.0 is an original implementation of the DES algorithm and the
crypt(3) interfaces used in Unix-style operating systems. It was produced
in Australia and as such is not covered by U.S. export restrictions (at
least for copies that remain outside the U.S.).
History
=======
An earlier version of the FreeSec library was built using the UFC-crypt
package that is distributed as part of the GNU library. UFC-crypt did not
support the des_cipher() or des_setkey() functions, nor the new-style
crypt with long keys. These were implemented in FreeSec 0.2, but at least
one bug remained, where encryption would only succeed if either the salt
or the plaintext was zero. Because of its heritage FreeSec 0.2 was covered
by the GNU Library Licence.
FreeSec 1.0 is an original implementation by myself, and has been tested
against the verification suite I'd been using with FreeSec 0.2 (this is not
encumbered by any licence). FreeSec 1.0 is covered by a Berkeley-style
licence, which better fits into the *BSD hierarchy than the earlier GNU
licence.
Why should you use FreeSec?
===========================
FreeSec is intended as a replacement for the U.S.-only NetBSD libcrypt,
to act as a baseline for encryption functionality.
Some other packages (such as Eric Young's libdes package) are faster and
more complete than FreeSec, but typically have different licencing
arrangements. While some applications will justify the use of these
packages, the idea here is that everyone should have access to *at least*
the functionality of FreeSec.
Performance of FreeSec 1.0
==========================
I compare below the performance of three libcrypt implementations. As can be
seen, it's between the U.S. library and UFC-crypt. While the performance of
FreeSec 1.0 is good enough to keep me happy for now, I hope to improve it in
future versions. I was interested to note that while UFC-crypt is faster on
a 386, hardware characteristics can have markedly different effects on each
implementation.
386DX40, 128k cache | U.S. BSD | FreeSec 1.0 | FreeSec 0.2
CFLAGS=-O2 | | |
========================+===============+===============+==================
crypt (alternate keys) | 317 | 341 | 395
crypt/sec | | |
------------------------+---------------+---------------+------------------
crypt (constant key) | 317 | 368 | 436
crypt/sec | | |
------------------------+---------------+---------------+------------------
des_cipher( , , , 1) | 6037 | 7459 | 3343
blocks/sec | | |
------------------------+---------------+---------------+------------------
des_cipher( , , , 25) | 8871 | 9627 | 15926
blocks/sec | | |
Notes: The results tabled here are the average over 10 runs.
The entry/exit code for FreeSec 0.2's des_cipher() is particularly
inefficient, thus the anomalous result for single encryptions.
As an experiment using a machine with a larger register set and an
obscenely fast CPU, I obtained the following results:
60 MHz R4400 | FreeSec 1.0 | FreeSec 0.2
========================+=================================
crypt (alternate keys) | 2545 | 2702
crypt/sec | |
------------------------+---------------------------------
crypt (constant key) | 2852 | 2981
crypt/sec | |
------------------------+---------------------------------
des_cipher( , , , 1) | 56443 | 21409
blocks/sec | |
------------------------+---------------------------------
des_cipher( , , , 25) | 82531 | 18276
blocks/sec | |
Obviously your mileage will vary with your hardware and your compiler...
