Re: [Ocf-linux-users] No performances improvment with AMD Geode

SourceForge Headquarters 1320 Columbia Street Suite 310 San Diego, CA 92101 +1 (858) 422-6466

Jivin "Adam Cécile (Le_Vert)" lays it down ...
> David McCullough a écrit :
> >Jivin "Adam Cécile (Le_Vert)" lays it down ...
> >...
> >  
> >>gandalf@wrap:~$ cat /etc/modules
> >># /etc/modules: kernel modules to load at boot time.
> >>#
> >># This file contains the names of kernel modules that should be loaded
> >># at boot time, one per line. Lines beginning with "#" are ignored.
> >>
> >># Geode WatchDog
> >>geodewdt
> >>
> >># Geode LX AES hardware crypto
> >>geode-aes
> >>geode-rng
> >>crypto_blkcipher
> >>ocf
> >>cryptosoft
> >>cryptodev
> >>
> >># Alix LEDs driver
> >>leds-alix
> >>
> >># Sensors
> >>lm90
> >>
> >>wrap:/home/gandalf# openssl speed -evp aes128 -elapsed
> >>You have chosen to measure elapsed time instead of user CPU time.
> >>To get the most accurate results, try to run this
> >>program when this computer is idle.
> >>Doing aes-128-cbc for 3s on 16 size blocks: 104083 aes-128-cbc's in 3.00s
> >>Doing aes-128-cbc for 3s on 64 size blocks: 101925 aes-128-cbc's in 3.00s
> >>Doing aes-128-cbc for 3s on 256 size blocks: 89677 aes-128-cbc's in 3.00s
> >>Doing aes-128-cbc for 3s on 1024 size blocks: 61529 aes-128-cbc's in 3.00s
> >>Doing aes-128-cbc for 3s on 2048 size blocks: 42862 aes-128-cbc's in 3.00s
> >>OpenSSL 0.9.8g 19 Oct 2007
> >>built on: Wed Jul  9 18:31:21 UTC 2008
> >>options:bn(64,32) md2(int) rc4(idx,int) des(ptr,risc1,16,long) 
> >>aes(partial) blowfish(idx)
> >>compiler: gcc -fPIC -DOPENSSL_PIC -DZLIB -DOPENSSL_THREADS -D_REENTRANT 
> >>-DDSO_DLFCN -DHAVE_DLFCN_H -DHAVE_CRYPTODEV -DL_ENDIAN -DTERMIO -O3 
> >>-march=i586 -Wa,--noexecstack -g -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: ftime
> >>The 'numbers' are in 1000s of bytes per second processed.
> >>type             16 bytes     64 bytes    256 bytes   1024 bytes   2048 
> >>bytes
> >>aes-128-cbc        555.11k     2173.68k     7652.44k    21001.90k    
> >>29260.46k
> >>
> >>wrap:/home/gandalf# lsmod | grep cry
> >>crypto_blkcipher       18564  1 geode_aes
> >>wrap:/home/gandalf# lsmod | grep ocf
> >>wrap:/home/gandalf# openssl speed -evp aes128 -elapsed
> >>You have chosen to measure elapsed time instead of user CPU time.
> >>To get the most accurate results, try to run this
> >>program when this computer is idle.
> >>Doing aes-128-cbc for 3s on 16 size blocks: 895899 aes-128-cbc's in 3.00s
> >>Doing aes-128-cbc for 3s on 64 size blocks: 324390 aes-128-cbc's in 3.00s
> >>Doing aes-128-cbc for 3s on 256 size blocks: 92243 aes-128-cbc's in 3.00s
> >>Doing aes-128-cbc for 3s on 1024 size blocks: 23823 aes-128-cbc's in 3.00s
> >>Doing aes-128-cbc for 3s on 2048 size blocks: 11998 aes-128-cbc's in 3.00s
> >>OpenSSL 0.9.8g 19 Oct 2007
> >>built on: Wed Jul  9 18:31:21 UTC 2008
> >>options:bn(64,32) md2(int) rc4(idx,int) des(ptr,risc1,16,long) 
> >>aes(partial) blowfish(idx)
> >>compiler: gcc -fPIC -DOPENSSL_PIC -DZLIB -DOPENSSL_THREADS -D_REENTRANT 
> >>-DDSO_DLFCN -DHAVE_DLFCN_H -DHAVE_CRYPTODEV -DL_ENDIAN -DTERMIO -O3 
> >>-march=i586 -Wa,--noexecstack -g -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: ftime
> >>The 'numbers' are in 1000s of bytes per second processed.
> >>type             16 bytes     64 bytes    256 bytes   1024 bytes   2048 
> >>bytes
> >>aes-128-cbc       4778.13k     6920.32k     7871.40k     8131.58k     
> >>8190.63k
> >>
> >>Okay there's performance difference when I unload all ocf/cryptodev 
> >>kernel modules. But are you sure it comes from hardware ? It may just 
> >>come from the fact that ocf software perform better than the standard 
> >>linux crypto stack. What's your mind about this ?
> >>    
> >
> >If ocf is using software crypto it can never (on a single crypto stream)
> >go faster than openssl (or kernel crypto) without ocf.  It's just not
> >possible since they are all using basically the same crypto code on the
> >same speed CPU.
> >
> >So yes,  I am sure you are doing HW crypto using OCF+cryptosoft and geode.
> >
> >It would go even faster if you ran it from in the kernel.  This is what
> >the ocf-bench driver does if you are interested:
> >
> >	modprobe ocf-bench
> >	dmesg
> >
> >Cheers,
> >Davidm
> >  
> Hello,
> 
> I've just rebuild my kernel with ocf-bench. Here is what happens:
> gandalf@wrap:~$ sudo modprobe ocf-bench
> [sudo] password for root:
> FATAL: Error inserting ocf_bench 
> (/lib/modules/2.6.25-2-alix/kernel/crypto/ocf/o                             
> cf-bench.ko): Invalid argument
> gandalf@wrap:~$ dmesg | tail
> [   44.840932] wan: no IPv6 routers present
> [   61.510480] ip6_tables: (C) 2000-2006 Netfilter Core Team
> [   66.773388] warning: `named' uses 32-bit capabilities (legacy support 
> in use)
> [   74.301323] tun: Universal TUN/TAP device driver, 1.6
> [   74.301346] tun: (C) 1999-2004 Max Krasnyansky <ma...@qu...>
> [   74.307817] tun0: Disabled Privacy Extensions
> [   82.660103] wlan: no IPv6 routers present
> [  132.637766] Crypto Speed tests
> [  132.637766] OCF: testing ...
> [  133.811467] OCF: 1044 requests of 1500 bytes in 294 jiffies
> gandalf@wrap:~$
> 
> Seems the module fails to load. Known issue?

No,  it fails to load by design, but you can see there that is printed
out a performance number.  You will need to find out you jiffies and
multiply that out to see the performance.

By default ocf-bench tests sha1/3des,  you will need to modify the
source to try out AES or something else.  I know it's a complete hack,
but it is useful if you want raw kenerl performance.

Cheers,
Davidm

-- 
David McCullough,  dav...@se...,   Ph:+61 734352815
Secure Computing - SnapGear  http://www.uCdot.org   http://www.snapgear.com