Investigating this on HECToR, you can also do the following:
aprun -n NUMPROCS time executable
Which only times your executable, although it will produce an output for each executable run (so if you're using 48 cores then you'll get 48 results). So you should see a time that is much closer to you program time.
I suggest try one run like this and see what you get.
honestly I don't see much point: it's going to be difficult to parse multiple outputs and I've already proven that the overhead is static. What do you think?
The output is actually easy to read/parse, I would just try it with one experiment and see the results you get. It can confirm whether the static overhead is due to aprun or to your application, which would be beneficial to definitely confirm.
Investigating this on HECToR, you can also do the following:
ReplyDeleteaprun -n NUMPROCS time executable
Which only times your executable, although it will produce an output for each executable run (so if you're using 48 cores then you'll get 48 results). So you should see a time that is much closer to you program time.
I suggest try one run like this and see what you get.
cheers
adrianj
honestly I don't see much point: it's going to be difficult to parse multiple outputs and I've already proven that the overhead is static. What do you think?
ReplyDeleteKK
The output is actually easy to read/parse, I would just try it with one experiment and see the results you get. It can confirm whether the static overhead is due to aprun or to your application, which would be beneficial to definitely confirm.
ReplyDeleteadrianj
OK, I'll do it, I think I see how to parse it
ReplyDelete