[dm-crypt] Any practical AES-NI benchmarks around?

[dm-crypt] Any practical AES-NI benchmarks around?

[Zdenek Kaspar]

Hello, does anyone have AES-NI capable CPU and is able to share some
dm-crypt numbers ?

TIA, Z.

Re: [dm-crypt] Any practical AES-NI benchmarks around?

[Rick Moritz]

Those Sandra results sure look promising, I wonder why no "real world" benchmark mirrors the extreme performance gain of the raw encryption power - one would expect gains that reflect the order of magnitude higher bandwidth.
For me, the killer is still cost. In the enterprise the new 1156 Xeon's will probably make great sense, but putting down 200 euro/dollar for a dual core doesn't seem viable at the moment. Maybe my server upgrade, will have to wait for sandy bridge after all. (Also, there are no H5x based mainboards with sufficient SATA ports for my needs, which also is hindering my acceptance of the platform)

On Sat, 23 Jan 2010 00:56:28 +0100 Zdenek Kaspar <zkaspar82@gmail.com> wrote:

> Dne 23.1.2010 0:27, Rick Moritz napsal(a):
> > Hi,
> > 
> > I saw some rough data in the clarkdale review at anandtech, their AES benchmark algorithm saw an increase in performance of a few percent (http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=3704&p=5). Not as much as I was hoping for. Additionally, the feature is not available on the i3 series cpu's, another setback, as I was hoping on getting decent AES performance on the cheap..
> > If anyone finds a more relevant benchmark, I'd also like to see it. At the moment I don't even know if we have a kernel module that does AES-N1 yet.
> > 
> > 
> > On Sat, 23 Jan 2010 00:16:59 +0100 Zdenek Kaspar <zkaspar82@gmail.com> wrote:
> > 
> >> Hello, does anyone have AES-NI capable CPU and is able to share some
> >> dm-crypt numbers ?
> >>
> >> TIA, Z.
> 
> Hi Rick, I quickly saw this one (promising):
> http://www.tomshardware.com/reviews/intel-clarkdale-core-i5-661,2514-3.html
> 
> And there are patches for openssl integration which makes me think it
> will be useful in future. It got merged (aesni-intel) in .30 and
> extended in .31 kernel.
> 
> Yeah, the price for dualcore CPU is not interesting atm unless it can
> cure dm-crypt SMP scaling issue..
> 
> Z.

Re: [dm-crypt] Any practical AES-NI benchmarks around?

[Rick Moritz]

On Fri, 22 Jan 2010 19:46:05 -0800 Bryan Kadzban <cryptsetup@kdzbn.homelinux.net> wrote:

> Now, on a solid-state device, you might see a (maybe) 2x speedup overall
> with 10x-faster encryption.  (Again pulling numbers out of thin air.)
> But not on a rotating disk; encryption is not where the bottleneck is.


Re-reading the anandtech comparison, I realise that the clarkdale i5 is a dual core CPU, and compares favourably to a lower clocked quad core (though both run 4 threads and cost roughly the same), so your factor 2 number for SSDs seams reasonably close. Still, I expect that the inclusion on lower power cpu's would make more sense, as these high end chips (3.2ghz dual core with 4 threads is the smallest chip supporting AES-NI) already pack quite a punch, so do not truly necessite the extra performance - the only benefit is that encryption does become more rtansparent, and not the cpu-hog that it is now.
With the i5's all costing exactly the same, there's only one obvious reason to skip the 750: AES cycles/watt...
I was hoping for a CPU that I could passively cool, like my current Sempron, and yet get 10-20 times the AES performance. Seems that I've got to wait a while yet, if I want to stick with commodity hardware.
Well, let's hope for a module to crystalize into the mainline kernel soon, so that we can at least have our own numbers -seems a patch already exists.