[IPOL discuss] CPU info and parallel code [Re: about performance comparisons in IPOL articles]

[Nicolas Limare]

Hi,

> When you compare software performances in an IPOL article, please be
> precise. Saying that "program A is twice as fast as program B" is not
> verifiable. I think that, to be meaningful, one must at least provide
> this information:
> * exact version of the codes
> * exact compiler version and compilation options
> * harwdare model, at least the CPU (model and speed) and RAM (size and
>   speed)

Be careful when reporting the number of CPUs on your machine. Many
processors have the "HyperThreading Technology", which alows 2
(or more) process to be simultaneously managed by a computing unit. In
that case, the operationg system usually shows you two processors
instead of one, and manages this processor almost as if it was
duplicated.

But you only have the computing power of one processor. It can
efficiently switch between two processes, but it only does real
computation for one process at a time. Moreover, running a
parallelized program in this environemnt is worse than running it
unparallelized, because of the parallelizing and switching cost.

So, if you do some parallel computing, you should:

* Check if your CPU has HyperThreading enabled. In Linux, use the
  command `cat /proc/cpuinfo | grep flags` and look for the "ht" flag.
* Disable it; you can disable HyperThreading from the BIOS. It is
  disabled on IPOL   servers.
* Describe your hardware precisely in your articles: which exact CPU
  model, how many CPUs (rarely more than one on laptops), how many
  cores per CPU, how many threads per core.

-- 
Nicolas LIMARE - CMLA - ENS Cachan         http://limare.perso.math.cnrs.fr/
IPOL journal                                             http://www.ipol.im/
-> image processing, reproducible research, open science
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 198 bytes
Desc: Digital signature
URL: <https://tools.ipol.im/mailman/archive/discuss/attachments/20130620/dd5aa764/attachment.pgp>