Re: gp timers # and ## in parallel mode

[Jerome Raulin <jerome@raulinfoissac.fr>]

Hi Bill,
Thanks very much for the upgrade, it's better than having to choose between cpu time or wall time.
I used to prefer wall clock time and so used the --time=ftime switch when calling the Configure script. With this configuration the evolution broke the build in the init.c file, walltimer_start function. I had to remove the --time switch for a successful build. With this new timing the ftime option is unnecessary as both cpu and wall clock is provided, the ftime option should probably be deprecated.
Many thanks.
Jérôme.

On January 22, 2019 11:09:20 PM GMT+01:00, Bill Allombert <Bill.Allombert@math.u-bordeaux.fr> wrote:

Dear PARI developers,

I have update the display of the gp timers # and ## in parallel GP
to print both the CPU time and the real time. This give a more useful
information.

For example with mt=pthread with 4 threads

? #
   timer = 1 (on)
? sizebyte(polmodular(101))
cpu time = 9,296 ms, real time = 2,641 ms.
%1 = 5538416

In this case the CPU time is the total CPU time for all threads.
We see that the cpu time is about 4 time the real time. Since my 
laptop has only 2 physical cores, the cpu time is about twice the
value with pthread disabled.

Now with mt=mpi with 5 threads:

? sizebyte(polmodular(101))
%1 = 5538416
? ##
  ***   last result: cpu time 1,480 ms, real time 3,686 ms.

In this case the CPU time is only the CPU time of the main thread.
Since most of the work is done in the secondary threads, we see that
the CPU time is smaller than the real time in this case.

Cheers,
Bill