🚀 go-pugleaf

Dan.Pop@mail.cern.ch (Dan Pop) wrote:
>In <41dejm$d1p@sundog.tiac.net> Jim Williams <williams@tiac.net> writes:
>>Will you agree that 1) memory is slower than cpu, and 2) most
>>programs don't fit in cache?  If this is true then the external
>>bus is generally the bottleneck. 
>
>Not necessarily.  You can write programs where the external bus is the
>bottleneck, but this is not automatically true in the general case.
>I/O bound and cpu bound programs are at least as common as memory bound
>programs.

Someone else made similar points (less politely), but that message
doesn't seem to have made it to my news server...

I'm not so sure that I/O and cpu are directly related, but I must admit
that the newer methods of CPU load balancing probably make my statements
a bit off-base.  Either I should go back and study current chip
designs more carefully or bitness should be dropped as an indicator
of chip throughput.  I vote we do away with bitness. :)

>
>>I'm perfectly happy with a
>>definition of bitness which revolves about the bottleneck... 
>
>But you're obviously missing the fact that the bottleneck is application
>specific (and some applications can be affected by more than one
>bottleneck).  For someone generating fractals or doing data
>compression, the cpu is the bottleneck.  And guess where is the
>bottleneck for the "tar" and "cp" commands?  How about "ftp" and "rcp"? :-)
>

I didn't really miss the application specific nature of bottlenecks,
but I do seem off on the relationship between bitness as a measure
of throughput and the possible location of bottlenecks.  It was easier
in the old days when memory, I/O, and just about everything else was
out on the bus.

With the newer placements of a little here a little there I really
wonder if bitness is a meaningfull concept.  If you can't use it
to roughly gauge throughput then what good is it?

>>In
>>fact that seems like a very good definition to me -- the bitness
>>of a machine is the path width of that portion of of the machine
>>which happens to be the usual bottleneck.  
>
>Too bad your definition is rendered useless by the fact that there is no
>usual bottleneck.  When memory started to become a real bottleneck, the
>chip manufacturers introduced the built-in cache.
>

There must be some normal distibution of bottleneckness (nice word).
Where are the current bottlenecks, and what loads cause them?  If
I can't use bitness as a measure then I need to find a replacement.

-- 
Sphere.

Find a Linux/GNU Group for you: http://www.tiac.net/users/williams/lugnuts/
Buy Free UNIX!
Microsoft is prohibited from examining any packet containing data
originating on any machine which I am using. (Not that I can stop them.)