[M3devel] lock performance, random thoughts

Tony Hosking hosking at cs.purdue.edu
Tue Apr 6 01:45:20 CEST 2010


Yes, that's pretty much what modern Java implementations do.


On 5 Apr 2010, at 19:35, Dragiša Durić wrote:

> I've been thinking few days about this lock performance thing. As POSIX
> mutexes are getting us too many times into kernel space by way of
> syscalls, maybe we can get some performance back while not loosing
> scheduling efficiency and multi-processor side of kernel-space/POSIX
> threading?
> 
> User-space threads were two things. One - user-space threads. Two -
> user-space synchronizing. If we implement former inCritical by using
> lockfree abilities of modern multi-processor/multi-core/cluster systems
> and then revert to user-space synchronizing _only_ we can have best of
> both worlds. 
> 
> We need user-space threads because we need to have our Modula-3 process
> spread it's threads over whatever we run it on. And we will be very
> happy to have synchronizing efficency back.
> 
> Few details to decide how to implement would be preempting/wakeup cycle,
> and waiting loop on our inCritical flag. Both can be solved by POSIX
> primitives.
> 
> What do you think about this?
> 
> dd
> 
> -- 
> Dragiša Durić <dragisha at m3w.org>




More information about the M3devel mailing list