Re: [boost] [Boost.Interprocess] Chronic performance on Win7

Subject: Re: [boost] [Boost.Interprocess] Chronic performance on Win7
From: Ion Gaztañaga (igaztanaga_at_[hidden])
Date: 2011-08-19 19:39:34

El 19/08/2011 12:52, Gav Wood escribió:

> Comments?

Thanks for the report. I must admit that process-shared synchronization 
implemenation in windows is awful at best. I have some ideas to 
implement those using named mutexes or events, but the fact is that I 
don't know any open source that emulates PTHREAD_PROCESS_SHARED properly.

Sleep(0) is a problem because emulated mutexes can't implement Priority 
inheritance (only kernel sync mechanisms can do that). Sleep(1) gives a 
chance to lower priority threads. The spin count until 20ms depends on 
the processor speed, which is not nice. Gav Wood (thanks!) has proposed 
SwitchToThread but I read here that:

http://msdn.microsoft.com/en-us/library/ms686352%28v=vs.85%29.aspx

"Windows 2003 warning!
If you plan to use SwitchToThread inside a function that is going to be 
called heavily, and your program is going to run on Windows 2003 server 
(whatever flavor), you better be careful. SwitchToThread executes 
painfully slow on w2003. My benchmarks show that a program running under 
Windows XP can call a function that performs a useful job and has 
SwitchToTherad on it, 40 million times in less than 2 minutes, while the 
same process (running on the same hardware) will take almost 40 minutes 
on Windows 2003."

I don't know if this is applicable also to other Windows versions.

Could you measure SwitchToThread version?

Best,

Ion