Is Intel Atom 270 really so slow?

[Stuart Henderson]

On 2009-10-02, MK <public at kubikcz.net> wrote:
> Hello
>
> I've just installed OpenBSD 4.5 on Intel Atom 270 1.6GHz mini ITX board and 
> made some tests.
> CPU speed does not look very good, in fact I was comparing OpenSSL tests 
> with my old Pentium III 500MHz and it's in many cases faster.
> (in ubench Pentium III is just about 1200 points lower than 1.6Ghz Atom)
>
> e.g.
>
> openssl speed aes
>
> Intel Atom
>
> The 'numbers' are in 1000s of bytes per second processed.
> type             16 bytes     64 bytes    256 bytes   1024 bytes   8192 
> bytes
> aes-128 cbc      10396.85k    12298.42k    12928.09k    13091.55k 
> 13142.58k
> aes-192 cbc       9116.41k    10525.00k    10981.04k    11102.07k 
> 11139.49k
> aes-256 cbc       8068.51k     9088.34k     9410.93k     9491.90k 
> 9520.14k
>
> Pentium III
>
> The 'numbers' are in 1000s of bytes per second processed.
> type             16 bytes     64 bytes    256 bytes   1024 bytes   8192 
> bytes
> aes-128 cbc      14980.42k    15963.78k    16222.05k    16329.53k 
> 16919.33k
> aes-192 cbc      13102.40k    14225.10k    14473.31k    14457.72k 
> 14487.41k
> aes-256 cbc      11976.37k    12662.91k    12850.40k    12842.31k 
> 12945.35k
>
> Is that an expected result, or is it caused by some wrong configuration on 
> my system?

They are not fast CPUs. The PIII/Pentium M derived CPUs are much
faster clock-for-clock. Don't expect anything like the performance of
a 1.4GHz Pentium M from an Atom system.

But you are also not using the recommended way to get accurate speed
test results.

e.g. openssl speed -evp aes256 -elapsed, openssl speed -evp aes128 -elapsed
http://markmail.org/message/27kslswk4zahitit?q=thread:ngdhgyjfi2bgx3mb

Here are examples from some of my systems.

type             16 bytes     64 bytes    256 bytes   1024 bytes   8192 bytes

hw.model=Intel(R) Atom(TM) CPU 230 @ 1.60GHz ("GenuineIntel" 686-class)
aes-128-cbc      19038.03k    22711.28k    25477.35k    26380.26k    26533.04k
aes-256-cbc      12845.55k    17726.87k    18514.90k    18995.34k    19078.13k

hw.model=Intel(R) Pentium(R) M processor 1.80GHz ("GenuineIntel" 686-class)
*throttled to 600MHz*
aes-128-cbc       9899.51k    24677.09k    27413.80k    28302.58k    28499.52k
aes-256-cbc      14628.60k    18896.27k    19306.85k    20853.67k    20903.13k

hw.model=Intel(R) Pentium(R) M processor 1.80GHz ("GenuineIntel" 686-class)
aes-128-cbc      29764.09k    58372.77k    76936.57k    84923.22k    85870.56k
aes-256-cbc      43417.90k    56709.43k    61328.90k    62481.61k    62912.43k

hw.model=Intel(R) Core(TM)2 CPU T7200 @ 2.00GHz
aes-128-cbc      91020.19k    99403.47k   101908.04k   102832.04k   102777.22k
aes-256-cbc      72070.90k    76275.03k    77913.78k    78081.53k    80180.70k

hw.model=AMD Opteron(tm) Processor 252
aes-128-cbc     103124.01k   113282.39k   117289.34k   118376.37k   118693.44k
aes-256-cbc      83422.11k    91675.96k    94265.24k    94990.56k    95180.43k

You'll see interesting results on the Geode LX in the soekris 5501/Alix
systems. They have hardware acceleration for aes-128-cbc but not aes-256-cbc.
There are setup overheads which aren't related to how much data is encrypted
at once, so large block sizes produce much higher rates than small ones (and
software does better than hardware for small block sizes).

hw.model=Geode(TM) Integrated Processor by AMD PCS ("AuthenticAMD" 586-class)

sysctl kern.usercrypto=1 (the default):
aes-128-cbc [1]    706.55k     2639.08k     8397.03k    18902.58k    28729.13k
aes-256-cbc [1]    619.83k     1694.13k     2956.24k     3628.23k     3943.12k

sysctl kern.usercrypto=0:
aes-128-cbc [2]   5808.83k     7405.93k     7868.92k     8148.31k     8241.87k
aes-256-cbc [2]   3990.27k     5351.07k     5948.26k     6035.84k     6048.84k