openssl on 64bit is much slower than before

[cc]

Posting output of openssl speed on precise alpha and 10.04. Some quick googling seems to indicate issues with 64bit asm vs. the C version (might be a good idea to disable 64bit asm). This is on a processor *without* AES-NI (which is still very common).

Precise:
type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes
aes-128 cbc 61725.52k 67902.33k 69264.13k 146084.86k 148518.23k
aes-192 cbc 52222.95k 56592.77k 57542.04k 122330.11k 125277.32k
aes-256 cbc 44912.70k 48644.01k 49539.94k 105978.88k 107372.74k

Lucid:
type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes
aes-128 cbc 74876.87k 118355.41k 138862.76k 144140.50k 141562.78k
aes-192 cbc 82093.08k 110813.18k 120764.33k 118315.44k 118527.32k
aes-256 cbc 74760.65k 96842.01k 104520.45k 106615.81k 103481.34k

ProblemType: Bug
DistroRelease: Ubuntu 12.04
Package: openssl 1.0.0g-1ubuntu1
ProcVersionSignature: Ubuntu 3.2.0-12.21-generic 3.2.2
Uname: Linux 3.2.0-12-generic x86_64
ApportVersion: 1.93-0ubuntu2
Architecture: amd64
Date: Fri Feb 24 02:33:44 2012
EcryptfsInUse: Yes
InstallationMedia: Ubuntu-Server 12.04 LTS "Precise Pangolin" - Alpha amd64 (20120201.1)
ProcEnviron:
 TERM=xterm-256color
 LANG=en_US.UTF-8
 SHELL=/bin/bash
SourcePackage: openssl
UpgradeStatus: No upgrade log present (probably fresh install)

[cc (codecrumb)]

Here's a post that talks about the old assembly for 64bit x86.
http://zombe.es/post/4059999783/openssl-outmoded-asm

Performance is up to 50% worse for AES ciphers so I would think this is a must fix.

Patch available for RC4 on x86_64. http://cvs.openssl.org/filediff?f=openssl/crypto/rc4/asm/rc4-x86_64.pl&v1=1.12&v2=1.13

For the rest, the option should be to disable ASM for 64bit and rely on AESNI for processors with AES-NI support.
Possible patch here: http://dl.dropbox.com/u/24257718/openssl-1.0.0d-noasm_aes_rc4.patch

Another data point:
http://old.nabble.com/64-bit-OpenSSL-FIPS-1.2.3-with-asm-slow-performance-problems-on-AES-td32773268.html

[Scott Moser (smoser)]

Just a reference, I've seen something similar at bug 963420, but my initial tests have not shown significant difference in i386 compared to amd64.

[cc (codecrumb)]

Can you post some numbers to compare? Post the output of "openssl speed aes" for different scenarios at a minimum and specify your processor type.

There should be a comparison of both 32bit to 64bit on precise (which is what I think you are comparing) AND a comparison of ssl performance on precise and earlier distributions.

Here are the results of my latest round of testing, my hardy results are from a 32bit build:

http://pastebin.com/XrFUT8Zq

So looking at the results of aes-128 (which arguably is more important currently than -192 or -256 because of wider usage), results of 16 - 256 bytes are off significantly.

[cc (codecrumb)]

Any decision on this?

The stopgap is not that hard. Disable ASM for 64bit and retest to see the performance numbers.

[cc (codecrumb)]

Just checking in 3 months later. Any chance on getting some movement on this for the next Precise update?

[Launchpad Janitor (janitor)]

Status changed to 'Confirmed' because the bug affects multiple users.

[Oibaf (oibaf)]

Can you still reproduce this issue on a newer and supported Ubuntu release?

[Launchpad Janitor (janitor)]

[Expired for openssl (Ubuntu) because there has been no activity for 60 days.]