[771] Delay and Timer granularity
Locked
[771] Delay and Timer granularity
 
|
What is the expected duration of "(Delay forMilliseconds: 1) wait"? I
was surprised to find that it actually takes 15ms, even with the new Timer mechanism enabled, and on Windows>=Vista with UseThreadpoolAPI=true. Interestingly, it's not just an overhead, it's a rounding up to the next 15ms. Delays from 1-15ms all take 15ms, and from 16-30ms all take 30ms. The same results are seen with "Timer after: 1 milliseconds signal: sem". The documentation in TimerSystem says that I should see a granularity of ~1ms. Is that incorrect, or am I doing something wrong? I've attached my test code - run the first line, save and restart, then inspect the rest. Cheers, Steve _______________________________________________ vwnc mailing list [hidden email] http://lists.cs.uiuc.edu/mailman/listinfo/vwnc |
Delay&TimerGranularity.txt (624 bytes)
Re: [771] Delay and Timer granularity
|
Re: [771] Delay and Timer granularity
|
|
Henrik Hoyer wrote:
> This is normal behavior on Windows > > http://www.geisswerks.com/ryan/FAQS/timing.html The simpler timing mechanisms on NT-based Windows have a granularity of 15ms (or 10ms for single processor machines). But VW can use smarter mechanisms, e.g. the microsecond clock is now accurate, as seen in #microsecondsToRun:. And the documentation of the new native Timers says they should be around 1ms, as I wrote. Steve PS The link you reference is rather out of date, missing info on XP, Vista and Win7. Better information can be found e.g. here: http://stackoverflow.com/questions/3763544/what-periodic-timer-objects-a re-available-in-the-windows-api _______________________________________________ vwnc mailing list [hidden email] http://lists.cs.uiuc.edu/mailman/listinfo/vwnc |
Re: [771] Delay and Timer granularity
|
|
In reply to this post by Steven Kelly
Off the top of my head, I wonder what is the behavior of
QueryPerformanceCounter() on that machine. We had an instance of a customer for which (Delay forSeconds: 1) wait took ~20 seconds of wall time on an XP machine. As it turns out, MSDN states in various places that, on XP, you might need the /usepmtimer boot.ini switch to work around various bugs that have accumulated over the years. I'd go take a look at the QueryPerformanceCounter() MSDN documentation from a Vista point of view. On 9/23/2011 5:02 AM, Steven Kelly wrote: > What is the expected duration of "(Delay forMilliseconds: 1) wait"? I > was surprised to find that it actually takes 15ms, even with the new > Timer mechanism enabled, and on Windows>=Vista with > UseThreadpoolAPI=true. Interestingly, it's not just an overhead, it's a > rounding up to the next 15ms. Delays from 1-15ms all take 15ms, and from > 16-30ms all take 30ms. The same results are seen with "Timer after: 1 > milliseconds signal: sem". > > The documentation in TimerSystem says that I should see a granularity of > ~1ms. Is that incorrect, or am I doing something wrong? I've attached my > test code - run the first line, save and restart, then inspect the rest. > > Cheers, > Steve > > > > > _______________________________________________ > vwnc mailing list > [hidden email] > http://lists.cs.uiuc.edu/mailman/listinfo/vwnc vwnc mailing list [hidden email] http://lists.cs.uiuc.edu/mailman/listinfo/vwnc |
Re: [771] Delay and Timer granularity
|
|
Thanks Andres. It looks like the new threadpool timers might not work as well on Windows 7. I've now tried on three Windows 7 machines with different hardware and OS histories, and all gave the same results as below. The microsecond clock works fine, not sure if that uses QueryPerformanceCounter().
On an old Vista laptop, the results were much better, around the documented ~1ms granularity and accuracy.
Since our Windows customers are on XP and Windows 7, I guess I either need to expect 15ms granularity, or then roll my own for now. It would be nice to have a timer that works on all three Windows versions with a 1ms granularity - Timer Queues seem to offer one: http://msdn.microsoft.com/en-us/library/ms686796(v=VS.85).aspx
http://www.informit.com/guides/content.aspx?g=dotnet&seqNum=816 (see 815 & 817 too)
Steve
From: [hidden email] on behalf of Andres Valloud Sent: Sat 24/09/2011 09:34 To: [hidden email] Subject: Re: [vwnc] [771] Delay and Timer granularity Off the top of my head, I wonder what is the behavior of _______________________________________________ vwnc mailing list [hidden email] http://lists.cs.uiuc.edu/mailman/listinfo/vwnc |
Re: [771] Delay and Timer granularity
|
|
Just tried again on the Vista machine after hibernate+wake, and now that too gives 15ms. Starting the same programs that were open during the previous test doesn't change this. I guess some other program had asked for 1ms timer granularity (with timeBeginPeriod?). Maybe VW should offer that too? Off by default, since it reduces battery life by 25%. From: Steven Kelly Thanks Andres. It looks like the new threadpool timers might not work as well on Windows 7. I've now tried on three Windows 7 machines with different hardware and OS histories, and all gave the same results as below. The microsecond clock works fine, not sure if that uses QueryPerformanceCounter().
On an old Vista laptop, the results were much better, around the documented ~1ms granularity and accuracy.
Since our Windows customers are on XP and Windows 7, I guess I either need to expect 15ms granularity, or then roll my own for now. It would be nice to have a timer that works on all three Windows versions with a 1ms granularity - Timer Queues seem to offer one: http://msdn.microsoft.com/en-us/library/ms686796(v=VS.85).aspx
http://www.informit.com/guides/content.aspx?g=dotnet&seqNum=816 (see 815 & 817 too)
Steve
From: [hidden email] on behalf of Andres Valloud Sent: Sat 24/09/2011 09:34 To: [hidden email] Subject: Re: [vwnc] [771] Delay and Timer granularity Off the top of my head, I wonder what is the behavior of _______________________________________________ vwnc mailing list [hidden email] http://lists.cs.uiuc.edu/mailman/listinfo/vwnc |
Re: [771] Delay and Timer granularity
|
|
And the solution is... (drum-roll)... open a Youtube video! If that's in a visible tab of my browser, on Win7 I get 1ms granularity. If it's in a non-visible tab, I get 10ms granularity. From: Steven Kelly Just tried again on the Vista machine after hibernate+wake, and now that too gives 15ms. Starting the same programs that were open during the previous test doesn't change this. I guess some other program had asked for 1ms timer granularity (with timeBeginPeriod?). Maybe VW should offer that too? Off by default, since it reduces battery life by 25%. From: Steven Kelly Thanks Andres. It looks like the new threadpool timers might not work as well on Windows 7. I've now tried on three Windows 7 machines with different hardware and OS histories, and all gave the same results as below. The microsecond clock works fine, not sure if that uses QueryPerformanceCounter().
On an old Vista laptop, the results were much better, around the documented ~1ms granularity and accuracy.
Since our Windows customers are on XP and Windows 7, I guess I either need to expect 15ms granularity, or then roll my own for now. It would be nice to have a timer that works on all three Windows versions with a 1ms granularity - Timer Queues seem to offer one: http://msdn.microsoft.com/en-us/library/ms686796(v=VS.85).aspx
http://www.informit.com/guides/content.aspx?g=dotnet&seqNum=816 (see 815 & 817 too)
Steve
From: [hidden email] on behalf of Andres Valloud Sent: Sat 24/09/2011 09:34 To: [hidden email] Subject: Re: [vwnc] [771] Delay and Timer granularity Off the top of my head, I wonder what is the behavior of _______________________________________________ vwnc mailing list [hidden email] http://lists.cs.uiuc.edu/mailman/listinfo/vwnc |
Re: [771] Delay and Timer granularity
|
|
Hi Steve, We use WaitForSingleObject in the "classic" timers (we manage our own sorted queue of timer alarms instead of having the operating system do it). However, we never call timeBeginPeriod(). If you'd be so kind as to stop watching youtube videos; could you please try calling timeBeginPeriod() from DLLCC in your image to see if it can affect your timer resolution? timeBeginPeriod: resolution <C: unsigned int timeBeginPeriod(unsigned int resolution)> ^self externalAccessFailedWith: _errorCode If the result is > 0 then it failed and you can see what kind of error here: http://pinvoke.net/default.aspx/winmm/MMRESULT.html This could be a very interesting change, as timeBeginPeriod should be transparent to all other win32 calls we make from the VM. Cheers, Michael On Sep 24, 2011, at 10:26 AM, Steven Kelly wrote:
_______________________________________________ vwnc mailing list [hidden email] http://lists.cs.uiuc.edu/mailman/listinfo/vwnc |
Re: [771] Delay and Timer granularity
|
|
Hi Michael, Fortunately I could watch the Youtube videos on my other computer, so my day wasn’t ruined ;). The test is a success, timeBeginPeriod works for both classic and native timers and delays. With the classic TimerSystem, timers and delays have an “overhead” of about 1ms; with the native TimerSystem, there is no noticeable overhead. #classicTimer -> #(1896 2990 3988 4889 5981 6977 7977 8973 9970 10967 11964 12962 13958 14955 15851 16950 17944 18939 19941 20941) #classicDelay -> #(1991 2892 3988 4986 5882 6977 7977 8973 9970 10965 11964 12962 13956 14955 15953 16949 17945 18943 19944 20933) #nativeTimer -> #(1004 2000 4085 4208 4993 5993 6984 7984 8981 9973 10981 11975 12973 13964 14972 15964 16963 17961 18958 19961) #nativeDelay -> #(1069 2007 3005 3989 5086 5992 6980 7986 8974 9988 10978 11976 12971 13969 14970 15966 16960 17963 18962 19957) Windows 7 SP1 64-bit, Intel i5, VW 7.7.1 Version Id: #[68 47 69 160 77 1 0 0 68 47 69 160] I’ve published WindowsTimerResolution in the public store repository. To improve the resolution, call #milliseconds:while:, e.g. WindowsTimerResolution milliseconds: 1 while: [Time microsecondsToRun: [1000 timesRepeat: [(Delay forMilliseconds: 1) wait]]] Cheers, Steve From: Michael Lucas-Smith [mailto:[hidden email]] Hi Steve, We use WaitForSingleObject in the "classic" timers (we manage our own sorted queue of timer alarms instead of having the operating system do it). However, we never call timeBeginPeriod(). If you'd be so kind as to stop watching youtube videos; could you please try calling timeBeginPeriod() from DLLCC in your image to see if it can affect your timer resolution? timeBeginPeriod: resolution <C: unsigned int timeBeginPeriod(unsigned int resolution)> ^self externalAccessFailedWith: _errorCode If the result is > 0 then it failed and you can see what kind of error here: http://pinvoke.net/default.aspx/winmm/MMRESULT.html This could be a very interesting change, as timeBeginPeriod should be transparent to all other win32 calls we make from the VM. Cheers, Michael On Sep 24, 2011, at 10:26 AM, Steven Kelly wrote: And the solution is... (drum-roll)... open a Youtube video! If that's in a visible tab of my browser, on Win7 I get 1ms granularity. If it's in a non-visible tab, I get 10ms granularity. From: Steven Kelly Just tried again on the Vista machine after hibernate+wake, and now that too gives 15ms. Starting the same programs that were open during the previous test doesn't change this. I guess some other program had asked for 1ms timer granularity (with timeBeginPeriod?). Maybe VW should offer that too? Off by default, since it reduces battery life by 25%. From: Steven Kelly Thanks Andres. It looks like the new threadpool timers might not work as well on Windows 7. I've now tried on three Windows 7 machines with different hardware and OS histories, and all gave the same results as below. The microsecond clock works fine, not sure if that uses QueryPerformanceCounter(). On an old Vista laptop, the results were much better, around the documented ~1ms granularity and accuracy. Since our Windows customers are on XP and Windows 7, I guess I either need to expect 15ms granularity, or then roll my own for now. It would be nice to have a timer that works on all three Windows versions with a 1ms granularity - Timer Queues seem to offer one: http://msdn.microsoft.com/en-us/library/ms686796(v=VS.85).aspx http://www.informit.com/guides/content.aspx?g=dotnet&seqNum=816 (see 815 & 817 too) Steve From: [hidden email] on behalf of Andres Valloud Off the top of my head, I wonder what is the behavior of _______________________________________________ _______________________________________________ vwnc mailing list [hidden email] http://lists.cs.uiuc.edu/mailman/listinfo/vwnc |
| Free forum by Nabble | Edit this page |