Details
-
Bug
-
Resolution: Delivered
-
P4
-
8
-
x86
-
windows_7
Description
FULL PRODUCT VERSION :
1.8.0-b132 32 bit (running as client VM)
FULL OS VERSION :
Windows 7 SP1 latest version 2014-04-14
EXTRA RELEVANT SYSTEM CONFIGURATION :
Older Intel Core i5-650 Westmere processor with "aes" flag enabled.
A DESCRIPTION OF THE PROBLEM :
AES-NI is clearly not being utilized on the i5-650, even in a 1000 x loop using two buffers (input/output) of 1 MiB. Neither works with the -XX:+UseAES -XX:+UseAESIntrinsics flags (though these seem to be the default). Not a millisecond difference with -XX:-UseAES -XX:-UseAESIntrinsics.
Tested using the same version Hotspot on x64 machine (Intel i7 3770) using 64 bit server VM, did work on that machine.
Results:
-AES on 650: 21 seconds
+AES on 650: 21 seconds
-AES on 3770: 8 seconds
+AES (or default) on 3770: 3 seconds
According to the info in "What's new"
http://www.oracle.com/technetwork/java/javase/8-whats-new-2157071.html
"The hardware must be 2010 or newer Westmere hardware. ".
So the i5-650 should be the first processor to support AES-NI intrinsics, right?
STEPS TO FOLLOW TO REPRODUCE THE PROBLEM :
Run AES in CBC mode using NoPadding in a loop using the SunJCE provider.
EXPECTED VERSUS ACTUAL BEHAVIOR :
Expected: faster than without AES-NI
Actual: same speed, so no AES-NI or exactly the same speed (unlikely and too slow for that)
REPRODUCIBILITY :
This bug can be reproduced always.
---------- BEGIN SOURCE ----------
// TODO
---------- END SOURCE ----------
CUSTOMER SUBMITTED WORKAROUND :
Use Netscape NSS in the PKCS#11 provider for AES encryption.
1.8.0-b132 32 bit (running as client VM)
FULL OS VERSION :
Windows 7 SP1 latest version 2014-04-14
EXTRA RELEVANT SYSTEM CONFIGURATION :
Older Intel Core i5-650 Westmere processor with "aes" flag enabled.
A DESCRIPTION OF THE PROBLEM :
AES-NI is clearly not being utilized on the i5-650, even in a 1000 x loop using two buffers (input/output) of 1 MiB. Neither works with the -XX:+UseAES -XX:+UseAESIntrinsics flags (though these seem to be the default). Not a millisecond difference with -XX:-UseAES -XX:-UseAESIntrinsics.
Tested using the same version Hotspot on x64 machine (Intel i7 3770) using 64 bit server VM, did work on that machine.
Results:
-AES on 650: 21 seconds
+AES on 650: 21 seconds
-AES on 3770: 8 seconds
+AES (or default) on 3770: 3 seconds
According to the info in "What's new"
http://www.oracle.com/technetwork/java/javase/8-whats-new-2157071.html
"The hardware must be 2010 or newer Westmere hardware. ".
So the i5-650 should be the first processor to support AES-NI intrinsics, right?
STEPS TO FOLLOW TO REPRODUCE THE PROBLEM :
Run AES in CBC mode using NoPadding in a loop using the SunJCE provider.
EXPECTED VERSUS ACTUAL BEHAVIOR :
Expected: faster than without AES-NI
Actual: same speed, so no AES-NI or exactly the same speed (unlikely and too slow for that)
REPRODUCIBILITY :
This bug can be reproduced always.
---------- BEGIN SOURCE ----------
// TODO
---------- END SOURCE ----------
CUSTOMER SUBMITTED WORKAROUND :
Use Netscape NSS in the PKCS#11 provider for AES encryption.