RFR: 8040140 System.nanoTime() is slow and non-monotonic on OS X
staffan.larsen at oracle.com
Tue Apr 15 09:20:05 UTC 2014
On 15 apr 2014, at 09:14, David Holmes <david.holmes at oracle.com> wrote:
> Hi Staffan,
> Generally looks okay.
> os_bsd.cpp still shows the old URL for Dave Dice's article
I had forgotten to save the file. :-(
> In the Solaris changes there is a lot of old code with inaccurate comments, but I suppose cleaning that up (oldgetTimeNanos()) is out of scope. You only added the check for AssumeMonotonicOSTimers in the supports_cx8 path, but the other path is now dead code.
Since supports_cx8() returns true (because SUPPORTS_NATIVE_CX8 is defined) on both solaris sparc and solaris x86, it looks like oldgetTimeNanos() is really dead code. I can remove it as part of this change if that’s ok.
> Do we need to document this only affects OSX and Solaris? (Though implicitly this acts as-if true on Linux and Windows in the common case.)
I will update the description to:
experimental(bool, AssumeMonotonicOSTimers, false, \
"Assume that the OS system timers are monotonic " \
"(Solaris and OS X)") \
> #ifdef TARGET_OS_FAMILY_bsd
> # include "jvm_bsd.h"
> # include <setjmp.h>
> + # include <mach/mach_time.h>
> I think this include needs to be in a OSX/Apple specific conditional.
Changed it to:
# include "jvm_bsd.h"
# include <setjmp.h>
# ifdef __APPLE__
# include <mach/mach_time.h>
> We should really fix the non-monotonic-clock path in the Linux and Windows implementations too ... but 32-bit is problematic <sigh>
> On 15/04/2014 4:00 PM, Staffan Larsen wrote:
>> Here is an updated webrev with changes to the comments in os_bsd.cpp and
>> - obs -> obsv
>> - fixed URL to blog entry
>> On 15 apr 2014, at 07:52, Staffan Larsen <staffan.larsen at oracle.com
>> <mailto:staffan.larsen at oracle.com>> wrote:
>>> On 14 apr 2014, at 21:08, Aleksey Shipilev
>>> <aleksey.shipilev at oracle.com <mailto:aleksey.shipilev at oracle.com>> wrote:
>>>> On 04/14/2014 06:55 PM, Staffan Larsen wrote:
>>>>> mach_absolute_time() is essentially a direct call to RDTSC, but with
>>>>> conversion factor to offset for any system sleeps and frequency
>>>>> changes. The call returns something that can be converted to
>>>>> nanoseconds using information from mach_timebase_info(). Calls to
>>>>> mach_absolute_time() do not enter the kernel and are very fast. The
>>>>> resulting time has nanosecond precision and as good accuracy as one
>>>>> can get.
>>>> Some numbers would be good on the public list :) I know the numbers
>>>> already, but others on this list don’t.
>>> I posted the numbers in the bug, but forgot to say so here...
>>>>> Since the value from RDTSC can be subject to drifting between CPUs,
>>>>> we implement safeguards for this to make sure we never return a lower
>>>>> value than the previous values. This adds some overhead to nanoTime()
>>>>> but guards us against possible bugs in the OS. For users who are
>>>>> willing to trust the OS and need the fastest possible calls to
>>>>> System.nanoTime(), we add a flag to disable this safeguard:
>>>> I now wonder if this safeguard can produce a stream of exactly the same
>>>> timestamps if local clock is lagging behind. But considering the
>>>> alternative of answering the retrograde time, and the observation the
>>>> current Mac OS X mach_absolute_time() *appears* monotonic, having this
>>>> safeguard seems OK.
>>>>> webrev: http://cr.openjdk.java.net/~sla/8040140/webrev.00/
>>>>> bug: https://bugs.openjdk.java.net/browse/JDK-8040140
>>>> This looks good to me.
>>>> And, since this question will inevitably pop up, do we plan to bring it
>>>> into 8uX? I think many Mac users will be happy about that.
>>> I would like to do so, but I would also like to have it sit and bake
>>> for a while in 9 before that. I think the 8u20 train has left the
>>> station, but perhaps 8u40?
More information about the hotspot-runtime-dev