Re: What would you like in DF 1.8?

[Matthew Dillon <dillon@xxxxxxxxxxxxxxxxxxxx>]

:It would be nice if the developers tried to 
:optimize the pthread implementation significantly.
:Moreover it would be awesome if the developers 
:could add some features in the kernel which 
:help the system provide all its horsepower for
:scientific computing applications. Fot that 
:we need maximum throughput and cache utilization 
:as well as great scalability of parallel codes.
:
:I am not a OS specialist but I can assure you that
:a lot of people from scientific computing society
:are disappointed with current operating systems 
:because they are not tweaked to provide the best 
:performance they can for numerically/memory intensive
:codes. 

    We do have an LWP (light weight processes) effort going on.  It won't
    make it into 1.6 but work has been progressing on it slowly.  We also
    have libthread_xu.  That threading library is primarily waiting
    on a new signals implementation and LWP to really become useful.

:Furthermore parallel performance of OpenMP codes
:running on Linux are miles behind that of AIX.
:I wonder what happens with BSDs. 
:
:Do you find the idea of providing kernel features 
:which tweak the system for running UP/SMP scientific 
:computing applications, interesting?
:
:Cheers!
:Simeon.

    Ultimately I think the only way to make those sorts of applications
    optimal is with an OS that is able to cluster generically, which is
    our goal with DragonFly, and then to build more specific mechanisms
    on top of the generic clustering.

					-Matt
					Matthew Dillon 
					<dillon@xxxxxxxxxxxxx>