Mac OS X System.nanoTime Improvements

[David Holmes]

On 15/05/2013 5:27 PM, Staffan Larsen wrote:
>
> On 15 maj 2013, at 05:00, David Holmes <David.Holmes at oracle.com> wrote:
>
>> On 15/05/2013 6:44 AM, Денис Ушаков wrote:
>>> Hi,
>>>
>>> couple of days ago Aleksey Shipilev mentioned that System.nanoTime() is very slow on OS X and is non-monotonic. After some research I found that there's no clock_gettime() function in OS X and current implementation falls back to gettimeofday().
>>
>> Wow! If that is the case that is truly horrible and the OSX porters should have been aware of that and flagged it. :( A nanoTime() based on gettimeofday would not be monotonic if ntp adjustments are being made.
>>
>>> OS X provides a monotonic clock function mach_absolute_time(). According to documentation (https://developer.apple.com/library/mac/#qa/qa1398/_index.html) it can be converted to nanoseconds using results of mach_timebase_info().
>>
>> We need to know more about how this works, and in particular whether it is using the TSC and so is subject to the well known problems using the TSC as a "clock" base. The API sounds similar to that on Windows: QueryPerformanceCounter + QueryPerformanceFrequency.
>
> The source is browsable here:
>
> mach_absolute_time:
> http://www.opensource.apple.com/source/xnu/xnu-2050.22.13/osfmk/i386/rtclock.c

Thanks Staffan - looks like it is a completely naive implementation that 
doesn't recognize the frequency slew you can get, nor that the TSC is 
unsynchronized across multiple processors. Maybe they don't run with any 
kind of frequency scaling enabled? And maybe they use some kind of 
external synchronization driver (like AMD had for Windows?). But I'm 
somewhat doubtful that is the case.

David
-----

> _rtc_nanotime_read:
> http://www.opensource.apple.com/source/xnu/xnu-2050.22.13/osfmk/x86_64/machine_routines_asm.s
>
> PAL_RTC_NANOTIME_READ_FAST:
> http://www.opensource.apple.com/source/xnu/xnu-2050.22.13/osfmk/i386/rtclock_asm_native.h
>
>
> /Staffan
>
>
>>
>> Thanks,
>> David
>> -----
>>
>>> Here's draft patch that uses mach_absolute_time() to provide fast and monotonic System.nanoTime() for OS X:
>>> https://gist.github.com/denofevil/db57fa547510b2dd87fd
>>>
>>> You can find benchmark written by Aleksey Shipilev here: https://github.com/shipilev/jmh-nanotime.
>>> Benchmark results with system specs: https://gist.github.com/denofevil/5026294ee6de42699b41
>>>
>>> Thanks,
>>> Dennis
>>>
>