[rescue] O2 rescue: advice for an SGI (relative) newbie

[Joshua Boyd]

On Tue, Mar 01, 2005 at 11:51:25PM -0600, Jonathan C. Patschke wrote:

> Be aware that just because you have an R10000 CPU in there doesn't make
> the O2 a fully 64-bit machine.  You can't run 64-bit binaries because
> the address bus is only 32 bits wide.  You can run n32 binaries which
> give you 64-bit arithmetic, though, which is where most performance
> gains from 64-bit code would be, anyway.

I'm not sure why the physical memory bus width is relevent.  I'm not
disputing whether the O2 has that n32 limitation though.  After all, do
UltraSparcs of the U1, U2, etc vintage have a physical address bus
greater than 32bits?  Using a larger physical address bus seems like a
waste if the machine isn't meant to hold more than 4 gigs of ram.
 
> 512MB is, indeed, quite comfortable.  That 18GB disk should also be
> plenty for IRIX.

I found 4 to be a pain.  9 is livable in the O2 but I was running into
space issues with 9 on the Octane.  The difference being that I
installed far more freeware stuff (particularly GCC) on the Octane than
I have on the O2.
 

> I understand that the alignment and padding problems have been fixed in
> GCC 3.recent, but I'm not convinced the compiler optimizes that well
> yet.  To be fair, I haven't side-by-side compared MIPS assembly code
> generated from each compiler since GCC 2.95.3, so it's quite possible
> that GCC is good stuff these days.

http://forums.nekochan.net/viewtopic.php?t=4966

There is a thread on nekochan about it.  There is a R10k patch for GCC
3.4 (maybe it exists for older versions, and maybe it has made it into
the 4.0 CVS tree).  The short summary is that a guy trying to run
assorted benchmarks is having trouble making MIPSpro perform better than
GCC. 
 
The thread got off to a bad start.  At one point one person says a stock
GCC is about 20% slower on his particular application than mipspro, gcc
with -O2 while the mipspro was -Ofast=ip35.  Not a very fair run for
GCC. 

http://forums.nekochan.net/viewtopic.php?p=37785#37785

The initiator finally runs a more even comparison.  Using atlas
benchmarks, GCC gets 523Mflops on the test system, while MIPSpro gets
525Mflops.  This lags about 5-8% behind what SGI's SCSL provides using
hand optimized code.  The target system in this case is a R12k 300mhz
Octane.  

Anyway, it makes GCC look more reasonable that it was in the past.
Still, I prefer to use mipspro if for no other reason than the better
error messages.

-- 
Joshua D. Boyd
jdboyd at jdboyd.net
http://www.jdboyd.net/
http://www.joshuaboyd.org/