Re: aes performance

[Arjan <n6bc23cpcduw@xxxxxxxxxxxxxxxxxxxxxx>]

Olaf Selke wrote:
> hello there,
> 
> as I understood tor spends most of its cpu time within openssl library aes crypto.
> Which result of "openssl speed aes" applies to tor? Is it aes-128 cbc 16 bytes?

It would be nice if Tor was using bigger blocks, but I've not looked at
the code yet.


> In this case my old Prestonia P4 Netburst Xeon box's throughput is supposed to
> be roughly about 40 MBit/s as middleman. Correct?
> 
> type             16 bytes     64 bytes    256 bytes   1024 bytes   8192 bytes
> aes-128 cbc      84098.99k   119729.69k   138053.97k   142741.16k   144386.04k
> aes-192 cbc      75035.35k   104143.72k   115681.81k   120099.84k   120949.42k
> aes-256 cbc      69559.47k    92221.78k   102006.05k   105361.75k   100274.74k
> 
> Strange to say that my desktop Core2 Duo E8400 @home performs only 33% better in
> openssl aes crypto than one of the old P4 Netburst Xeon cores from my tor node.
> For the sake of better performance I'm thinking about replacing my tor node's
> hardware.

If you're going to replace hardware, hardware assisted encryption may be
an option. Recent VIA CPUs like the C7 and the Nano can do that. Their
clock frequency isn't very high, so something else (instead of
encryption) may become the bottleneck.

Resuls for VIA Nano (with 32-bit openssl):

VIA Nano 1.6 GHz with padlock engine
type         16 bytes    64 bytes   256 bytes  1024 bytes  8192 bytes
aes-128-cbc  69796.09k  275407.68k  585574.57k  815018.33k  920136.36k
aes-192-cbc  52670.82k  208539.14k  472485.55k  691277.82k  798340.44k
aes-256-cbc  50984.77k  201934.27k  439964.25k  623764.14k  709612.89k

VIA Nano 1.6 GHz without padlock engine
type         16 bytes    64 bytes   256 bytes  1024 bytes  8192 bytes
aes-128-cbc  41429.86k   55836.29k   60886.87k   62508.03k   62974.63k
aes-192-cbc  35838.02k   47521.62k   51671.72k   52854.10k   53177.00k
aes-256-cbc  33208.77k   41789.16k   45009.83k   45891.24k   46153.73k

Performance for 16-byte blocks is pretty poor, but for bigger blocks
it's much faster.

A VIA C7 running at the same clock frequency should produce similar results.