phoronix test, openssl bad performance

phoronix test, openssl bad performance

[Xose Vazquez Perez]

hi,

In the openssl test, Solaris is 2.4X times faster than Linux.

The hardware was:
ThinkPad T61 with Intel Core 2 Duo T9300 + 2GB RAM

OS (all 32-bit):
Build 127 of OpenSolaris 2010.02 with 5.11 kernel + GCC 4.3.2
Fedora 12 Linux 2.6.31 kernel + GCC 4.4.2
Ubuntu 9.10 Linux 2.6.31 kernel + GCC 4.4.1

"All operating systems were left with their default options during
the installation process and left with the default set of packages
for each operating system except for the obvious need to install
some dependencies for the different tests."

In Linux, it means AppArmor/SELinux and audit enabled, and etc...

the test:
openssl-0.9.8k, compiled with ./config no-zlib; make
command: $ openssl speed rsa4096 -multi $NUM_CPU_CORES


results: http://www.phoronix.com/data/img/results/freebsd8_benchmarks/20.png

Solaris     72.67 sign/s
Linux     ~=30    sign/s
(FreeBSD    20.3  sign/s)


complete article at http://www.phoronix.com/scan.php?page=article&item=freebsd8_benchmarks

-thanks-

Re: phoronix test, openssl bad performance

[Andi Kleen]

Xose Vazquez Perez <xose.vazquez@gmail.com> writes:

> hi,
>
> In the openssl test, Solaris is 2.4X times faster than Linux.
>
> The hardware was:
> ThinkPad T61 with Intel Core 2 Duo T9300 + 2GB RAM
>
> OS (all 32-bit):
> Build 127 of OpenSolaris 2010.02 with 5.11 kernel + GCC 4.3.2
> Fedora 12 Linux 2.6.31 kernel + GCC 4.4.2
> Ubuntu 9.10 Linux 2.6.31 kernel + GCC 4.4.1

You should probably report that to Fedora through their bugzilla, 
it's unlikely to be a linux kernel problem.

-Andi

Re: phoronix test, openssl bad performance

[Xose Vazquez Perez]

On 12/10/2009 11:03 PM, Andi Kleen wrote:

> You should probably report that to Fedora through their bugzilla, 
> it's unlikely to be a linux kernel problem.

I got *same* results with Fedora kernel as with a _clean_ 2.6.32-kernel.org
Not a Fedora-kernel problem!

Re: phoronix test, openssl bad performance

[Eric Dumazet]

Le 11/12/2009 00:44, Xose Vazquez Perez a écrit :
> On 12/10/2009 11:03 PM, Andi Kleen wrote:
> 
>> You should probably report that to Fedora through their bugzilla, 
>> it's unlikely to be a linux kernel problem.
> 
> I got *same* results with Fedora kernel as with a _clean_ 2.6.32-kernel.org
> Not a Fedora-kernel problem!

Not a kernel problem anyway, this is a 100% user land workload

Only conclusion is that openssl was not correctly compiled on linux distros,
while/or better compiled on solaris (or its a 64 bit solaris, or test results wrong)


32bit kernel

first build :

time openssl speed rsa4096 -multi 2
...

OpenSSL 0.9.7a Feb 19 2003
built on: Fri Jan 12 09:26:34 EST 2007
options:bn(64,32) md2(int) rc4(idx,int) des(ptr,risc1,16,long) aes(partial) blowfish(idx)
compiler: gcc -fPIC -DZLIB -DOPENSSL_THREADS -D_REENTRANT -DDSO_DLFCN -DHAVE_DLFCN_H -DKRB5_MIT -DOPENSSL_NO_IDEA -DOPENSSL_NO_MDC2 -DOPENSSL_NO_RC5 -DOPENSSL_NO_EC -I/usr/kerberos/include -DL_ENDIAN -DTERMIO -Wall -O2 -g -pipe -m32 -march=i686 -mtune=pentium4 -Wa,--noexecstack -DSHA1_ASM -DMD5_ASM -DRMD160_ASM
available timing options: TIMES TIMEB HZ=100 [sysconf value]
timing function used:
                  sign    verify    sign/s verify/s
rsa 4096 bits 0.041368s 0.000614s     24.2   1628.0


perf report :

# Samples: 1141271
#
# Overhead  Command             Shared Object  Symbol
# ........  .......  ........................  ......
#
    66.53%  openssl  /lib/libcrypto.so.0.9.7a  [.] bn_mul_add_words
     9.08%  openssl  /lib/libcrypto.so.0.9.7a  [.] bn_sqr_comba8
     7.55%  openssl  /lib/libcrypto.so.0.9.7a  [.] bn_add_words
     5.55%  openssl  /lib/libcrypto.so.0.9.7a  [.] bn_sub_words
     2.21%  openssl  /lib/libcrypto.so.0.9.7a  [.] BN_from_montgomery
     2.19%  openssl  /lib/libcrypto.so.0.9.7a  [.] bn_mul_comba8
     1.91%  openssl  /lib/libcrypto.so.0.9.7a  [.] bn_sqr_recursive
     1.14%  openssl  /lib/libcrypto.so.0.9.7a  [.] OPENSSL_cleanse
     0.83%  openssl  /lib/libcrypto.so.0.9.7a  [.] BN_usub
     0.59%  openssl  /lib/libcrypto.so.0.9.7a  [.] BN_copy
     0.38%  openssl  /lib/libcrypto.so.0.9.7a  [.] bn_mul_recursive
     0.29%  openssl  /lib/libcrypto.so.0.9.7a  [.] bn_cmp_words
     0.25%  openssl  /lib/libcrypto.so.0.9.7a  [.] BN_mod_exp_mont_consttime
     0.20%  openssl  /lib/libcrypto.so.0.9.7a  [.] 0x00000000044cb6
     0.12%  openssl  /lib/libcrypto.so.0.9.7a  [.] 0x0000000002b496
     0.11%  openssl  /lib/libcrypto.so.0.9.7a  [.] bn_mul_words
     0.07%  openssl  /lib/libcrypto.so.0.9.7a  [.] BN_bin2bn
     0.06%  openssl  /lib/libcrypto.so.0.9.7a  [.] BN_CTX_start
     0.06%  openssl  /lib/libcrypto.so.0.9.7a  [.] BN_mod_mul_montgomery
     0.06%  openssl  /lib/libcrypto.so.0.9.7a  [.] BN_sqr
     0.05%  openssl  /lib/libcrypto.so.0.9.7a  [.] BN_CTX_get
     0.04%  openssl  /lib/libcrypto.so.0.9.7a  [.] BN_CTX_end
     0.04%  openssl  /lib/libcrypto.so.0.9.7a  [.] BN_bn2bin
     0.04%     perf  [kernel]                  [k] __copy_from_user_ll_nozero
     0.03%  openssl  [kernel]                  [k] _spin_lock
     0.03%  openssl  /lib/tls/libc-2.3.4.so    [.] _int_malloc
     0.03%  openssl  [kernel]                  [k] ktime_get
     0.02%  openssl  /lib/libcrypto.so.0.9.7a  [.] BN_ucmp
     0.02%  openssl  [kernel]                  [k] update_wall_time
     0.02%  openssl  /lib/libcrypto.so.0.9.7a  [.] BN_mul


Other build (probably uses SSE2) :

OpenSSL 0.9.8k 25 Mar 2009
built on: Thu Oct 29 10:41:08 CET 2009
options:bn(64,32) md2(int) rc4(idx,int) des(ptr,risc1,16,long) aes(partial) idea(int) blowfish(idx)
compiler: gcc -fPIC -DOPENSSL_PIC -DOPENSSL_THREADS -D_REENTRANT -DDSO_DLFCN -DHAVE_DLFCN_H -DL_ENDIAN -DTERMIO -O3 -fomit-frame-pointer -Wall -DOPENSSL_BN_ASM_PART_WORDS -DOPENSSL_IA32_SSE2 -DSHA1_ASM -DMD5_ASM -DRMD160_ASM -DAES_ASM
available timing options: TIMES TIMEB HZ=100 [sysconf value]
timing function used:
                  sign    verify    sign/s verify/s
rsa 4096 bits 0.028285s 0.000380s     35.4   2631.6


Then, same hardware/machine with 64 bits kernel/openssl :

OpenSSL 0.9.8k 25 Mar 2009
built on: Thu Dec 10 22:02:14 CET 2009
options:bn(64,64) md2(int) rc4(1x,char) des(idx,cisc,16,int) aes(partial) idea(int) blowfish(ptr2)
compiler: gcc -DOPENSSL_THREADS -D_REENTRANT -DDSO_DLFCN -DHAVE_DLFCN_H -m64 -DL_ENDIAN -DTERMIO -O3 -Wall -DMD32_REG_T=int -DOPENSSL_BN_ASM_MONT -DSHA1_ASM -DSHA256_ASM -DSHA512_ASM -DMD5_ASM -DAES_ASM
available timing options: TIMES TIMEB HZ=100 [sysconf value]
timing function used:
                  sign    verify    sign/s verify/s
rsa 4096 bits 0.017143s 0.000272s     58.3   3673.1

Re: phoronix test, openssl bad performance

[Andi Kleen]

On Fri, Dec 11, 2009 at 12:44:55AM +0100, Xose Vazquez Perez wrote:
> On 12/10/2009 11:03 PM, Andi Kleen wrote:
> 
> > You should probably report that to Fedora through their bugzilla, 
> > it's unlikely to be a linux kernel problem.
> 
> I got *same* results with Fedora kernel as with a _clean_ 2.6.32-kernel.org
> Not a Fedora-kernel problem!

What I meant is that the kernel should not normally affect SSL performance.
OpenSSL (or whatever SSL package is used) is all user-space. So it needs 
to be looked at whoever maintains the SSL libraries first.

linux-kernel is really only for kernel problems.

-Andi

-- 
ak@linux.intel.com -- Speaking for myself only.