Re: A 32bit/64bit performance data point
Oops, I originally replied only to Ron.
Well, it's the libgmp (Gnu Multi-Precission math library) folks and
the OpenSSL folks that you hope are generating efficient AMD64 code
for discrete mathematics over finite fields (DH, El Gamal, DSA) and
finite rings (RSA). GMP uses hand-optimized assembly on most
architectures, as does OpenSSL.
OpenSSL uses it's own big integer code (in crypto/bn). It looks like
for 4096-bit moduli, AMD64 asm runs 2 to 4 times as fast as IA-32
asm. (The note says IA-32 asm is about as fast as AMD64 C for the big
integer code.) Note that DSA for all key sizes still uses some
160-bit finite field math.
-Karl
from: http://cvs.openssl.org/getfile/openssl/crypto/bn/asm/x86_64-gcc.c?v=1.1.4.3
* Q. How much faster does it get?
* A. 'apps/openssl speed rsa dsa' output with no-asm:
*
* sign verify sign/s verify/s
* rsa 512 bits 0.0006s 0.0001s 1683.8 18456.2
* rsa 1024 bits 0.0028s 0.0002s 356.0 6407.0
* rsa 2048 bits 0.0172s 0.0005s 58.0 1957.8
* rsa 4096 bits 0.1155s 0.0018s 8.7 555.6
* sign verify sign/s verify/s
* dsa 512 bits 0.0005s 0.0006s 2100.8 1768.3
* dsa 1024 bits 0.0014s 0.0018s 692.3 559.2
* dsa 2048 bits 0.0049s 0.0061s 204.7 165.0
*
* 'apps/openssl speed rsa dsa' output with this module:
*
* sign verify sign/s verify/s
* rsa 512 bits 0.0004s 0.0000s 2767.1 33297.9
* rsa 1024 bits 0.0012s 0.0001s 867.4 14674.7
* rsa 2048 bits 0.0061s 0.0002s 164.0 5270.0
* rsa 4096 bits 0.0384s 0.0006s 26.1 1650.8
* sign verify sign/s verify/s
* dsa 512 bits 0.0002s 0.0003s 4442.2 3786.3
* dsa 1024 bits 0.0005s 0.0007s 1835.1 1497.4
* dsa 2048 bits 0.0016s 0.0020s 620.4 504.6
*
* For the reference. IA-32 assembler implementation performs
* very much like 64-bit code compiled with no-asm on the same
* machine.
*/
On Fri, 10 Dec 2004 07:40:16 -0600, Ron Johnson <ron.l.johnson@cox.net> wrote:
> On Fri, 2004-12-10 at 13:36 +0100, Bill Allombert wrote:
> [snip]
>
> > Also amd64 is the only 64bit extension I know that have a hardware
> > 64bitx64bit-to-128bit multiplication. Running this bench on sparc64
>
> How much would help crypto work?
>
> (I hope the GCC AMD64 people noticed, and generate the proper
> code... :)
>
> --
> -----------------------------------------------------------------
> Ron Johnson, Jr.
> Jefferson, LA USA
> PGP Key ID 8834C06B I prefer encrypted mail.
>
> "Vegetarians eat Vegetables, Humanitarians frighten me."
> Unknown
>
>
>
Reply to: