/ Zope / Apsis / Pound Mailing List / Archive / 2004 / 2004-08 / Pound+OpenSSL+AEP is slower than Pound+OpenSSL

[ << ] [ >> ]

[ pound-current oddity / "Jay West" ... ] [ Latest Pound and OWA (Exchange 2003) are now ... ]

Pound+OpenSSL+AEP is slower than Pound+OpenSSL
Motoharu Kubo <mkubo(at)3ware.co.jp>
2004-08-08 11:55:02 [ FULL ]
I have tested Pound with AEP cryptocard (AEP1000L) and got strange result
that software solution is faster than hardware solution.

First of all, OpenSSL 0.9.6m [engine] recognizes the card.   "openssl
speed rsa dsa" gave following results:

with "-engine aep"
                  sign    verify    sign/s verify/s
rsa  512 bits   0.0002s   0.0000s   5021.1  69571.4
rsa 1024 bits   0.0006s   0.0001s   1792.9  14805.1
rsa 2048 bits   0.0013s   0.0001s    770.6  10400.0
rsa 4096 bits   0.0053s   0.0084s    188.9    119.4
                  sign    verify    sign/s verify/s
dsa  512 bits   0.0004s   0.0005s   2569.1   1926.3
dsa 1024 bits   0.0004s   0.0005s   2302.9   1895.2

without "-engine aep"
                  sign    verify    sign/s verify/s
rsa  512 bits   0.0028s   0.0003s    362.3   3812.2
rsa 1024 bits   0.0126s   0.0007s     79.3   1427.6
rsa 2048 bits   0.0720s   0.0022s     13.9    458.6
rsa 4096 bits   0.4695s   0.0074s      2.1    135.8
                  sign    verify    sign/s verify/s
dsa  512 bits   0.0022s   0.0026s    452.1    379.6
dsa 1024 bits   0.0061s   0.0075s    163.8    133.1

With AEP card, OpenSSL can handle encryption/decription more than ten
times faster than without AEP card.

However, "http_load -rate 100 -seconds 10" test gives following results.
https://hostname/apache_pb.gif is
used as request URL.

with "SSLEngine aep"

281 fetches, 7 max parallel, 653606 bytes, in 10.0277 seconds
2326 mean bytes/connection
28.0223 fetches/sec, 65179.9 bytes/sec
msecs/connect: 113.087 mean, 174.391 max, 0.438 min
msecs/first-response: 100.485 mean, 174.391 max, 40.329 min

without "SSLEngine aep"

435 fetches, 9 max parallel, 1.01181e+06 bytes, in 10.0153 seconds
2326 mean bytes/connection
43.4338 fetches/sec, 101027 bytes/sec
msecs/connect: 75.62 mean, 147.836 max, 0.44 min
msecs/first-response: 73.6562 mean, 203.61 max, 28.849 min

I tried several times against several other URL but got similar results.

Could someone give me suggestion to improve performance of Pound used with
AEP card?, thank you.

Motoharu Kubo
mkubo(at)3ware.co.jp

Re: Pound+OpenSSL+AEP is slower than Pound+OpenSSL
"Yoshinori TAKESAKO" <y_takesako(at)dreamarts.co.jp>
2004-08-09 04:37:22 [ FULL ]
Hi, 

I think that the benchmark of "openssl speed" has not returned 
the correct value.

You have to use "-elapsed" option, or use OpenSSL version 0.9.7 series.
Because the OpenSSL-engine version 0.9.6 benchmark command can not 
deal correctly with the test in the ability of a hardware accelerator 
by the default. 

Would you answer about the following three questions? 

(1) What is the processor? (You are using)
    e.g. Pentium-III 1.13GHz 

(2) What is OS? (You are using)
    e.g. Red Hat Enterprise Linux AS3.0

(3) Please let me know the result which executed the following commands, 
    with "-engine aep" and without "-engine aep".
  # openssl speed -elapsed rsa dsa

At Sun, 08 Aug 2004 18:55:02 +0900
 Motoharu Kubo <mkubo(at)3ware.co.jp> wrote:[...]

I think that this result is strange...
Recently, AEP card is not so fast. (???)

The following URL is written in Japanese. (not English)
http://itpro.nikkeibp.co.jp/linux/backnum/200212/correct.shtml

Regards,


--
  Yoshinori TAKESAKO <y_takesako at dreamarts.co.jp>

Re: Pound+OpenSSL+AEP is slower than Pound+OpenSSL
Motoharu Kubo <mkubo(at)3ware.co.jp>
2004-08-09 05:49:27 [ FULL ]
Yoshinori, thanks for your help.
[...]

Celeron (Coppermine) 600MHz
[...]

Z-Linux, which is basically equal to Turbolinux 8 server in terms of 
kernel version, glibc, openssl etc.
[...]

with "-engine aep"
                   sign    verify    sign/s verify/s
rsa  512 bits   0.0045s   0.0017s    222.3    574.6
rsa 1024 bits   0.0139s   0.0028s     72.0    360.5
rsa 2048 bits   0.0722s   0.0045s     13.9    224.3
rsa 4096 bits   0.4647s   0.0084s      2.2    119.2
                   sign    verify    sign/s verify/s
dsa  512 bits   0.0043s   0.0080s    232.5    125.0
dsa 1024 bits   0.0082s   0.0159s    121.5     62.7

without "-engine aep"
                   sign    verify    sign/s verify/s
rsa  512 bits   0.0028s   0.0003s    362.0   3810.3
rsa 1024 bits   0.0126s   0.0007s     79.4   1442.1
rsa 2048 bits   0.0723s   0.0022s     13.8    460.8
rsa 4096 bits   0.6018s   0.0106s      1.7     93.9
                   sign    verify    sign/s verify/s
dsa  512 bits   0.0022s   0.0027s    445.3    369.3
dsa 1024 bits   0.0061s   0.0075s    164.3    133.5

Hmmmmmmm....., "with -engine aep" is slower than "without -engine aep" !!
I measured three times, but results are similar to above list.
[...][...]

Agreed.  The advantage of AEP card seems to reduce CPU usage.
[...]

For those who are not familiar with Japanese, the itpro.nikkeibp.co.jp 
page says:

   By default, benchmark program of OpenSSL-engine 0.9.6g can not deal with
   hardware accelerator.  To get correct result, -elapsed option with
   0.9.6g or OpenSSL 0.9.7 is needed....
[...]

Re: Pound+OpenSSL+AEP is slower than Pound+OpenSSL
"Yoshinori TAKESAKO" <y_takesako(at)dreamarts.co.jp>
2004-08-09 07:46:07 [ FULL ]
Hi, 

At Mon, 09 Aug 2004 12:49:27 +0900
 Motoharu Kubo <mkubo(at)3ware.co.jp> wrote:[...]

Celeron 600MHz... Since it is a little old machine, 
it may be that the PCI bus etc are bottlenecks, isn't it?
I think that AEP card required PCI V2.2(32bit/64bit) interface.
What is PCI version of the motherboard which you are using?
[...]

I regret it, but this result is as expected.
[...]

I think so, too.

How about asking for information to development company 
(http://www.aepsystems.com/) or AEP
distributors?

Regards,


--
  Yoshinori TAKESAKO <y_takesako at dreamarts.co.jp>

MailBoxer