Thank you David,
This is what I wanted to know and your answer is what I wanted to hear.
Hi Simeon,
On 27 Aug 2009, at 08:54, Simeon Mitev wrote:
Nice Nicola,
Thank you for the good presentation!
I have got answers for some of my questions. What I am still
interested in is multi threading support in objC compiler and runtime.
While the server application will run on unix with 2 quad core
processors I will definitively need to use multi threading to get a
real benefit from the hardware. In C++ I can use also openMP
(http://openmp.org/wp/) in example. Thoughts ?
OpenMP is intended to allow things like loops to be parallelized across
threads. It defines a series of pragmas that the user can use to
provide hints to the compiler. If you are going to use it, make sure
you understand the characteristics of the target machine. It is very
easy to use OpenMP to make slower code than without using it if you do
not understand the target's memory architecture, cache coherency model,
and scheduling system.
Yes, the target machine is well known, it is based on Intel x64 quad
core processors, that's why I am thinking how to get the maximum from
the hardware.
Hyperthreading is Intel's brand name for their chip multithreading
implementation, which splits the execution units between two contexts.
You absolutely and categorically do not want to be using OpenMP to
parallelize loops like the one you gave as an example on a machine with
hyperthreading, because it is likely to have very few cache misses (the
compiler should be inserting prefetch instructions at the right points
in the loop) and so you will just be confusing the operating system
scheduler and starving one of the threads of execution units.
I will go through OpenMP's documentation once again. For me is more
important to generate native for the target machine code.
There are currently two Objective-C compilers that you can use with
GNUstep; gcc and clang/llvm (llvm-gcc only works on OS X). GCC supports
OpenMP, Clang does not. Clang supports a number of Objective-C 2
constructs (e.g. fast enumeration, declared properties), GCC does not.
OpenMP attempts to implicitly create threads for you. You can, of
course, explicitly create threads if you wish. At the lowest level, you
can use the POSIX theading APIs. GNUstep provides NSThread on top of
these, which gives a simpler abstraction, along with a number of support
classes, like NSLock, NSCondition, and so on. You can also use
Distributed Objects between processes or threads, which provides quite a
heavy-weight system for concurrency. If your program consists of
largely-independent parts, then you should consider splitting it into
different programs and using DO for communication. This makes it easy
to add fault tolerance (you can easily restart a crashed process; a
crashed thread will kill the whole program) and lets you scale to
multiple machines in a cluster in future.
EtoileThread is part of EtoileFoundation, and provides another
abstraction using higher-order messaging. You can read a bit more about
it here:
http://etoileos.com/news/archive/2008/07/29/0231/
David
Thank you for your help David. It looks like I will consider to use
GNUStep Base and environment for the server development, Google Protocol
Buffers as communication protocol and XCode (Apple, Objective C) for
client application development. The database will be either Firebird or
BerkeleyDB. I am still concidering whether to use universal relational
database or heavily customized embedded object oriented database engine.
// Simeon
_______________________________________________
Discuss-gnustep mailing list
[email protected]
http://lists.gnu.org/mailman/listinfo/discuss-gnustep
