🚀 go-pugleaf

Michael <mutk@iprimus.com.au> wrote:
>Ray wrote:
><snipp>
>> Comparing kernel compile times on two different systems is worthless unless
>> you both have exactly the same options selected.
>> 
>
>
>That *isn't* what I did, and *isn't* what Floyd did.

Fact.

And, I decided last night to do a little more testing, just to
round it all out.  I compiled a non-SMP kernel for my older box.
This box has an Asus P2B-DS motherboard with 512Mb of RAM and a
pair of 533Mhz Celerons in slotkits.  It has a SCSI drive and an
IDE drive.  The figures from hdparms show memory access is 98
MB/s, the IDE drive is 20 MB/s and the SCSI drive is 13 MB/s.
All tests were done on a basically quiet machine.

First, with an SMP kernel running, I compiled the kernel with
the source code on the SCSI disk, with the command

     make clean; time make bzImage

The best time was 7:27.  I then ran it with the command

     make clean; time make -j2 MAKE="make -j2" bzImage

The best time was 4:21.  (A ratio of 1.7).

Then I did those again, with the source code on the IDE drive.
I only ran each one time, with 7:24 for the single process make,
and 4:21 for the multi-process make.  Three seconds difference
is basically about the minimum noise from one compile to the
next...

Then I rebooted with the non-SMP kernel and ran the same
tests again.  7:37 was the best time for a single process
make, and 7:32 was the best a multi-process make.

With the source on the IDE drive, and the  non-SMP kernel,
a single process make was 7:35 and a multi-process make was
7:32.

These tests confirmed what I had previously stated that I would
expect to see (which was based on vague memories from doing
somewhat simular testing about a year ago on the system with
dual Athlon MP1600+ cpu's).

 * The dual cpu machine is slightly faster (2%) than a single
   cpu machine when doing a make that does not invoke multiple
   parallel processes.

 * Invoking a multi-process make on a single cpu machine
   decreased the time by only 1%.

 * Compiling source code from a disk system that is nearly twice
   as fast had almost no effect (1%) on compile time.  The job
   appears to be compute bound.

 * On this system SMP can provide at least a 1.7 times speed
   increase for compute bound multiple process jobs.


-- 
Floyd L. Davidson         <http://www.ptialaska.net/~floyd>
Ukpeagvik (Barrow, Alaska)                 floyd@barrow.com