Re: GNUstep Solaris versus Apple MacOSX (performance comparison)

[Richard Frith-Macdonald]

On Monday, December 23, 2002, at 01:27  pm, Andreas Hoeschler wrote:

> Hi all,
>
> we are just in the process of porting our business logic from Cocoa to 
> GNUstep. This question is probably not only GNUstep related but might 
> also depend on the used hardware. However, I would like to ask what 
> experiences others have made with the performance of applications 
> ported fom Cocoa to GNUstep
>
> We have a "Dual G4 533MHz 512 MByte" running MacOSX on the left side 
> and a "Netra X1 500MHz 736MByte" running Solaris 8 GNUstep on the 
> right. Compiling sources on the Solaris machine takes 4 times as long 
> as on the Mac. The application itself also runs slower on the Solaris 
> machine. I know the Netra is the cheapest Sparc machine you can get 
> costing only a fraction of the expenses for a Mac. However, since the 
> Netra is 64Bit and has more memory I expected both machines to at 
> least have the same performance. Where does this huge difference come 
> from?
>
> Now is this only hardware related - if yes, what are the essential 
> differences between the Mac and the Netra - or is this more software 
> related (GNUstep code, gcc,...)?

The recent gcc compilers are well know to be very slow, so that could 
be why compilation on solaris is so slow.

It's hard to tell really what's down to hardware or software...

As a general rule though, I think you can say that the Apple ObjC 
runtime and memory management is faster than the GNU code, while the 
GNUstep base library is faster than Foundation, and the gui library is 
slower than AppKit.  Because the runtime and memory management are so 
fundamental, this means that overall the Apple stuff is faster than the 
GNU stuff.

That being said ... real program performance can vary HUGELY depending 
on what is being done ... for instance in a benchmark, use of the 
NSArray indexOfObjectIdenticalTo: method is nearly 400 times faster in 
GNUstep than in MacOS-X.  Doubtless similar disparities work against 
GNUstep too.
What this means is that you need to see how your program performs, find 
where the slow points are, and optimise.   The advantage of GNUstep 
here is that you are able to optimise the GNUstep code (and contribute 
back your optimisations) rather than just optimising your application 
code.  It's certainly possible to build programs that run faster under 
GNUstep than under MacOS-X, and with visible source, easier to see 
where any problems lie.

FYI  I enclose output from benchmark.m (in base/Testing) as run under 
MacOS-X and Debian on the same powerbook earlier today.  The Debian 
version was using a base library compiled for debug ... but that 
probably won't have slowed down the GNU benchmark significantly IMO.

MacOSbench
Description: Binary data

GNUbench
Description: Binary data