Re: openssl speed benchmarks (was Re: [Openvpn-users] a couple of questions)

[uml <um...@do...>]

Here are some quick benchmarks I did on various x86 hardware...

P4 2GHz
aes-128-cbc     33183.77k    36574.07k    38054.66k    38431.17k

Dual AMD XP1800+
(null)           77402.95k    83313.81k    85872.75k    86574.58k
86517.76k

Virtually:
OpenBSD 3.5 on VMWare GSX (single proc) (with 3 other vm's running
concurrently in the background)
Host = 2 x 2.8GHz Xeons (533MHz fsb)
aes-128-cbc      28882.25k    28809.11k    29936.46k    30044.44k
29917.48k

SuSE 9.1 using 2.4.26-um1 UML kernel on the same dual XP1800+ above: (single
thread)
aes-128-cbc      37723.93k    41557.39k    42516.57k    43189.25k
43291.99k

SuSE 9.1 using 2.6.6-um1 UML kernel on the same dual XP1800+ above: (single
thread)
aes-128-cbc      36038.45k    41595.88k    42864.98k    42881.37k
43100.84k

And once again, the dual XP1800+ but with a single thread:
aes-128-cbc      38804.26k    41933.97k    43069.95k    43424.06k
43540.48k

Interestingly, the UML kernels didn't drop much performance when running a
single thread compared to the host, I expected much more, but I suppose
that's for a different list. (=

----- Original Message ----- 
From: "David Haworth" <da...@fy...>
To: "OpenVPN Users List" <ope...@li...>
Sent: Wednesday, June 16, 2004 12.07 PM
Subject: Re: openssl speed benchmarks (was Re: [Openvpn-users] a couple of
questions)


> > a intel 3.06 GHz though not completely idle got:
> > aes-128 cbc      30871.89k    31260.69k    31642.28k    31615.66k
> > 31645.70k
> >
> > my powerbook 1.33Ghz got
> > aes-128 cbc      38118.21k    41516.88k    41268.42k    40523.62k
> > 41802.05k
>
> now that's interesting. the 1.3G g4 powerbook got a significantly
> higher score than a 3G p4. I wonder why. bus speed is lower, altivec
> perhaps? would openssl support that?
>
> > This is interesting if you compare it to your G5, which is a dual and
> > pr. cpu is about 35%
> > faster than my single G4. However if you look at your results, they
> > are just about 2.5
> > times faster. If you take the cpu speed factor, and the 2 cpus, that
> > gives 1.35*2 = 2,7
>
> well, 2.5 vs 2.7 seems fair enough, you rarely get twice the speed out
> of a duallie. that difference in performance would seem to imply to me
> that openssl scales very well over cpu's but that there is a little
> overhead from whatever source (osx perhaps) to balance things over both
> cpu's properly. certainly it would be interesting trying it on a quad
> and 8 way box too, not that I have one to hand :)
>
> I don't think that any of the improvements in the g5 proc are likely to
> make too much of a difference to openssl speeds though. there is no
> need for >32bit addressing and crypto is perhaps not the best type of
> program for long pipelines where branch prediction might cause a bigger
> problem (I'm clutching at straws here, not sure) so I really don't
> know.
>
> my g5 running only 1 thread gets this result
>
> aes-128-cbc      53196.23k    56806.46k    62308.81k    60952.53k
> 62393.86k
>
> which is slightly more than half the dual thread attempt, so that seems
> to be saying good things about openssl's threading.
>
> what I find very interesting is that using the athlon single proc.
> using multiple threads makes performance slightly worse...
> understandable. using more than 2 threads on the duallie gets very
> little advance. again, I understand this. but using multiple threads on
> the little via chip reaps big rewards, esp with the bigger block sizes,
> although it murders the machine with massive load averages and very
> laggy response.
>
> this is at 10 threads:
> (null)           24912.90k   107973.05k   303573.21k   606096.96k
> 1196668.44k
>
> 100 threads:
> (null)           37495.22k   161566.05k   439774.14k   783542.34k
> 1864134.09k
>
> ie I'd say that ethernet and pci bus traffic are the bottleneck to one
> of these cpu's
>
> without the evp flag it gets a score similar to your fw:
>
> aes-128 cbc       5709.79k     5792.67k     5814.92k     5797.15k
> 5831.65k
>
> incidentally, the aes acceleration can cope with higher rates than just
> 128bit
>
> aes-256-cbc      25523.18k    95334.00k   259990.24k   453591.16k
> 581554.67k
>
> the board I'm using is one of these:
>
> http://www.viavpsd.com/product/epia_MII_spec.jsp?motherboardId=202
> http://linitx.com/product_info.php?
> products_id=414&osCsid=918d32a8a21e5e1c7a3696d8907913a8
>
> apparently via are going to be adding sha1 and sha256 offload to their
> next processor, and some rsa help too. now if only they'd add blowfish
> for us geeks, and des&md5 cos damn, it would make a fantastic
> /etc/passwd cracking engine :)
>
> is that helpful?
>
> dave
>
>
>
> --
> David Haworth - da...@fy...
> www.fyonn.net - aim/ichat: fyonnuk
>