[aarch64-port-dev ] Some benchmarking help required

[Dmitry Chuyko]

On 05/16/2018 08:03 PM, Andrew Haley wrote:
> On 05/16/2018 03:48 PM, Dmitry Chuyko wrote:
>> On 05/16/2018 05:44 PM, Andrew Haley wrote:
>>> On 05/16/2018 03:21 PM, Dmitry Chuyko wrote:
>>>> I personally like the AES results on both machines! It looks like the
>>>> first short patch makes the difference.
>>> But that is really strange because we have AES hardware.  The C2
>>> compiler should not affect anything.
>> Yes. And both ones do have it.
> Here's the answer, from a run of crypto.aes:
>
> 7464759  40.0675  16257.jo                 java                     void com.sun.crypto.provider.DESedeCrypt.decryptBlock(byte[], int, byte[], int)~1
> 7461795  40.0516  16257.jo                 java                     int com.sun.crypto.provider.CipherBlockChaining.implEncrypt(byte[], int, int, byte[], int)~2
> 991198    5.3203  16257.jo                 java                     void com.sun.crypto.provider.DESCrypt.perm(int, int, byte[], int)
> 787277    4.2258  16257.jo                 java                     void com.sun.crypto.provider.AESCrypt.implDecryptBlock(byte[], int, byte[], int)
> 735359    3.9471  16257.jo                 java                     void com.sun.crypto.provider.AESCrypt.implEncryptBlock(byte[], int, byte[], int)
>
> Despite the name of the benchmark, it's doing a great deal of DES
> encryption.  The time taken by the AES code is relatively tiny.
>
In the code it does AES and DES equally on the same data with 2 padding 
options.

And here is the rest of the interesting data:

==T88

Benchmark               #1 %    p       #1#2 %  p
scimark.sor.small       3.07    0.041   3.13    0.007
scimark.sparse.small    -5.93   0.000   -5.24   0.000
xml.transform           16.76   0.000   14.58   0.000
xml.validation          2.72    0.005   0.25    0.398

==T99

Benchmark               #1 %    p       #1#2 %  p
scimark.lu.small        -6.04   0.041   0.06    0.474
scimark.sor.small       5.31    0.019   3.99    0.047
scimark.sparse.small    -1.18   0.000   -2.83   0.001
scimark.monte_carlo     8.36    0.000   11.56   0.000
serial                  4.24    0.001   2.21    0.051
xml.transform           4.87    0.000   4.61    0.000
xml.validation          4.82    0.000   5.40    0.000


Sparse-small -1-6% regression looks real. Monte Carlo on T99 looks 
promising as it may work so for many CAS loops. Other +3-17% numbers are 
also very nice.

-Dmitry