Next message: Steven G. Johnson: "Rules Clarifications: Timing"
I have two comments regarding the instructions for timing.
First, the last time I checked, clock() would not work
correctly for parallel programs written in Cilk because the
timers can be inconsistent between processors. Unless this
has been fixed, it would probably be a problem for any
multi-threaded program for which you can't be sure two
calls to clock() are being made on the same processor. If
anyone else is using Cilk or OS threads, I believe you
need to use gethrtime() (or Cilk_get_wall_time, which is
just an inline call to gethrtime). My information on this
may be out of date, so if anyone knows for sure, please
correct me.
Also, I am unclear on why we are dividing by the number of
processors when doing timing measurements. Shouldn't we
be multiplying the time by the number of processors if we
want MFLOPs per processor?
----
NOTE: This message was sent using a WWW form. The address rsschwar@mit.edu
was typed manually, and may easily be incorrect.
You may post a follow-up message or a new message.
To send a reply directly to the author, you
may click on the email address above.
If you would like to submit a message using
your own mail program, send it to:
18.337-wm@webmeet.mit.edu
If you are following up this article, please include the following
line at the beginning of your message:
In-Reply-To: 9704162028.AA21335@arachnophobia.MIT.EDU