[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: armel boxes for Debian



Am Donnerstag, 21. Februar 2008 09:18:35 schrieb Florian Weimer:
> Based on the numbers, I'd guess that the box hasn't got enough RAM.  Are
> you sure it's CPU-bound?
>
> > Based on these xulrunner build times, we'd need an arm box that's 4x as
> > fast as the current one to be "competitive":
> >
> > 	i386		0:42 (amd64 is of course approx. the same)
> > 	alpha		1:02
> > 	s390		1:10
> > 	sparc		1:59
> > 	ia64		1:19
> > 	hppa		2:39
> > 	powerpc		2:45
> > 	mips		3:09
> > 	mipsel		3:19
> > 	arm		13:01

It is definitely CPU-bound. Keep in mind, current arm cpus are neither 
superscalar nor do they have out-of-order execution. They can not even do 
loads in parallel to arithmetic instructions, which means a 600MHz Xscale is 
equivalent to an Athlon64 with 200...400MHz from terms of raw CPU power. If 
it comes to floating point (which is required for many script languages even 
if they deal with integer numbers), it is much worse, as we do not have a 
FPU.

But what is probably even more important for compilation, current arm cpus do 
not have L2 caches, so if data does not fit into the L1 cache (32kBytes in 
the XScale) they have to access the RAM. Unfortunately, the XScale's RAM 
interface is crippled by some serious design bugs. We only get a memcopy 
speed of around 40MB/s and a memory latency of 200ns. An Athlon64 can get a 
memcopy speed of 600MB/s and a latency of 50ns out of the same memory module, 
and if you put in two of them you get even a memcopy speed of 1130MB/s 
(latency does not improve of course).

If you take into account these facts, you must admit that the Thecus just does 
a good fight ;-)

Peter


Reply to: