MPI Zoo: message size classification
Sitemap
|
|
This is a follow-up of the first point of this previous post.
As we have seen previously, communication time is not linear with message size. Yet previous experiments did aggregate timing and measurements, which may bias results and hinders statistical analysis. Martin and Arnaud have set up an easy experimental framework based on conceptual (for describing the experiment to perform), akypuera (for tracing the experiment), paje (to visualize the corresponding trace), and R (for performing the analysis). conceptual makes a bunch of initial measurements at first to obtain information on the machines and synchronize clocks. We also run rastrotimesync before and after launching conceptual so as to possibly resynchronize traces. Akypuera traces are then merged into a paje trace that can be either visualized or converted to a simple text file easily readable by R using a simple perl script.
While I was in the train back home, I've done the analysis of two measurements Martin did just before I left. The analysis is in my Sweave document and provides interesting insights.
Entered on