[aarch64-port-dev ] Some benchmarking help required

[Dmitry Chuyko]

Hi Andrew,

I blindly applied first, second and both and tried SPECjvm2008 on T88 
and T99.

Runs are still in progress, I'll complete the data tomorrow. Some 
initial results are below:

1. With only patch #2 applied there was always SIGSEGV at least in 
mpegaudio, scimark.fft.large, scimark.lu.large and scimark.sor.large. I 
will check this and get back with more details.

2. T88. Difference by with p<0.05 (16 samples):

Benchmark		#1 %	p	#1#2 %	p
compress		-2.74	0.080	0.53	0.367
crypto.aes		*23.50*	0.000	*31.83*	0.000
crypto.rsa		-0.90	0.077	-2.11	0.000
derby			1.23	0.012	1.59	0.014
mpegaudio		4.04	0.000	3.08	0.000
scimark.fft.large	-3.03	0.001	1.07	0.094
scimark.lu.large	-2.53	0.011	-2.96	0.001
scimark.sor.large	4.77	0.038	10.60	0.000

3. T99. Difference by with p<0.05 (16 samples):

Benchmark		#1 %	p	#1#2 %	p
crypto.aes		*17.54*	0.000	*18.38*	0.000
crypto.rsa		2.01	0.000	1.22	0.000
crypto.signverify	0.76	0.000	1.09	0.000
derby			3.57	0.121	-8.56	0.000
mpegaudio		*17.55*	0.000	*17.23*	0.000
scimark.fft.large	0.35	0.038	-0.12	0.240


I personally like the AES results on both machines! It looks like the 
first short patch makes the difference.

-Dmitry

On 05/11/2018 07:37 PM, Andrew Haley wrote:
> I've noticed poor code generation for loops post-JDK 11.
>
> I've written a couple of patches to try to fix this, and I'd appreciate some
> help with testing them.  They are:
>
> http://cr.openjdk.java.net/~aph/aarch64_8203006-1.patch
> http://cr.openjdk.java.net/~aph/aarch64_8203006-2.patch
>
> I'd love someone to run SPEC jvm and SPEC jbb on both, preferably on more
> than one microarchitecture.  Thank you.
>