[rescue] Sigh

[Jonathan C. Patschke]

On Sat, 8 Jun 2002, Bill Bradford wrote:

> > But, yeah, I'd say an E6500 with slow CPUs is about comparable to a
> > challenege XL.  
> 
> Based on my experience with each of the above systems sitting in 
> $work's datacener, No Fucking Way.
> 
> Challenge XL *might* be as fast as a dual-300mhz U60.

Sorry.  I should've said PowerChallenge XL/R10000, which is what I was
thinking.  An original-run Challenge XL was bad-ass in its day, but it's
unfair to compare an R4400 to an UltraSPARC II.

The comparison really depends on the number/type of CPUs and what you're
trying to do.  I was thinking in terms of what I usually do to my
machines--relatively little I/O, more number/string-crunching.  The E6500
has much greater throughput across the midplane than Challenge XL, but I
think they're comparable for CPU-bound tasks (remember, I said -slow- CPUs
in the E6500--not 450MHz chips).

The 300MHz CPU in my U10 I have at home is about 15% faster than the
195MHz R10k in my Octane, unless I've done the math incorrectly.  This was
a fairly straightforward integer benchmark (DES encryption via a
standardized implementation of crypt() that -should- be small enough to
live in the cache).  Once you factor in that the Challenge doesn't have
XIO, and that the CPUs sit on a fairly classic SMP bus, there's a bit of
overhead to figure in.

So, yeah, a quad-processor Challenge XL is probably about comparable to a
dual-CPU U60, but that's a tiny configuration--almost no one bought a
refrigerator-sized SGI to just hold one CPU card.  A fully-loaded
Challenge XL at 36 195MHz CPUs will eat anything smaller than an E3000 for
lunch at strictly CPU-bound tasks--especially those that are easily
parallelized.  No one makes a dual-CPU system that fast yet.

> 6500 just *stomps* on everything else we've got.

E6500 has an utter assload of throughput to memory and periphreal I/O,
too--twice as much as PowerChallenge XL.  For databases and mail (which is
what those E6500s were labeled as doing), that's what you want.  
Challenge XL was alright in its day, but the 1994 limitations of its
architecture make a PowerChallenge XL/R10k really only suitable as a
compute server--not an I/O monster.

--Jonathan