The benchmark result looks good to me. I still have a few questions.
Read inlines, please.
On 5/24/2019 12:16 PM, Martin Balao wrote:
Hi Xuelei,
Thanks for your reply.
I think I now know what you mean.
Here it's a new benchmark:
http://cr.openjdk.java.net/~mbalao/webrevs/8223482/benchmark_sslcontextloading_manual_v0.tar.gz
In this new benchmark we measure the following sequence:
long startTime = System.currentTimeMillis();
ctx = SSLContext.getInstance("TLSv1.2");
long stopTime = System.currentTimeMillis();
The SSLContext class gets initialized per run.
We test both NON_FIPS (SunJCE) and FIPS (SunPKCS11) providers.
Results summary (100 runs per case):
FIPS_with_8223482_webrev01.txt average: 314.33 ms
NON_FIPS_with_8223482_webrev01.txt average: 817.91 ms
If I understand correctly, you run the test with the patch of webrev01?
http://cr.openjdk.java.net/~mbalao/webrevs/8223482/8223482.webrev.01/
FIPS_without_8223482_webrev01.txt average: 358.42 ms
NON_FIPS_without_8223482_webrev01.txt average: 771.34 ms
If I understand correctly, you run the test with the pacth of webrev00?
http://cr.openjdk.java.net/~mbalao/webrevs/8223482/8223482.webrev.00/
From the above numbers, the FIPS_with_8223482_webrev01 is better than
FIPS_without_8223482_webrev01, but NON_FIPS_with_8223482_webrev01 is
worse than NON_FIPS_without_8223482_webrev01. It is a little bit
strange to me.
Did you have the numbers for the latest JDK build, without any patch?
So, yes, it seems that there is a ~6% startup impact on SunJCE. These
numbers are not accurate though because of using
System.currentTimeMillis to measure, they provide just a rough idea.
A rough idea is okay. Maybe, you can use nanoseconds. But it is not
really necessary.
Thanks,
Xuelei
