comparing interpreter speed on two platforms

[Steven D. Majewski]

On 2 Oct 2001, Chris Liechti wrote:

> i have run a simple benchmark on two systems:
> 
> a standard PC:
> 1. PentiumPro 200Mhz (400BogoMips) Debian Linux 2.2 (Kernel 2.2-17) 64M RAM
> 
> and an embedded system:
> 2. PowerPC    ~60Mhz (68BogoMips)  HardHat Linux 2.0 (Kernel 2.4) 16MB RAM
> 

I suspect Cliff's diagnosis is the right one:

>Because, as stated in the Linux docs, BogoMips is a meaningless benchmark.

However, a couple of other possibilities: 

GCC isn't as well optimised for PPC as for x86: 
Apple has done some work on GCC but they had a separate fork until they
just recently merged it back for GCC 3.x. An earlier version probably
won't have any of apple's mods. 

math libs for PPC may be slow -- this was noticed on Darwin/MacOSX 
developer lists that the Darwin math libs were much much slower than
Metrowerks PPC math libs -- especially for trig functions. I don't 
know if those libs are part of GCC or BSD distribution, and I don't
know if the same problem exists on Linux, but if you have any sins or
cos's in your math benchmarks, I'ld guess this is a likely glitch. 

If you haven't already, you may want to break your benchmark up to
profile different aspects -- is it a general slowdown, or is one 
thing (like math libs, above) very slow? 

16MB vs. 64MB: it should all fit into 16MB easily, but is there a
difference in how quickly it pages in ?  Make sure you not only
have a minimum system load, but run an untimed pass first to make sure
everything has paged into memory. 


-- Steve Majewski