Duane Percox <[log in to unmask]> wrote:
> Yes - on a relative basis between similar systems running the
> same os you might be able to derive some reasonable assumptions.

Exactly.  Comparing hardware is only relevant if the *software* you
are going to be running on it is virtually identical.

As time goes on, the various parts of a system (CPU, disk capacity,
disk bandwidth, disk latency, memory capacity, memory bandwidth,
memory latency) improve at different rates.  Generally CPU speed
improves the fastest, and since the other components of the system
(not to mention the human end user) increase at a slower rate (or
not at all) there is "extra" CPU available, and people are pretty
good at using it up somehow.

Comparing the simple TurboImage to a fancy Oracle installation is a
good example where it might take 100,000 instructions to do in Oracle
what Image does in 1000.

If you want to compare raw CPU speed, then you need to know the "path
length" (number of instructions executed) for the task on each system
so you can appropriately scale the "MHz" into some kind of common
unit (actually you need a lot more information than just this to
do it well, but it's a useful illustration).

Another example: I have a benchmark application that compiles on a
1500MHz Itanium system with full optimization in about 18 seconds.
Profiling the compiler using HP's "caliper" tool shows me that the
compile executed ~30,000,000,000 (thirty billion) instructions in
the process.  The same program compiles on older HP-UX systems in
a bit longer time, but nowhere near as much longer as you might
expect if you work out that 30,000,000,000 instructions
would take around 7 minutes to execute on the older system (or
around two hours(!) on a classic 3000).  So clearly the IPF
compiler is taking advantage of its extra CPU power to perform
optimizations that would not have been practical to implement on
older, slower, systems and this makes it difficult to use it to
compare system speeds for example, just as will be the case between
any two software systems that might perform apparently the same
function but use different implementations.

G.

* To join/leave the list, search archives, change list settings, *
* etc., please visit http://raven.utc.edu/archives/hp3000-l.html *