Hello,
Happy New Year! I have some fresh benchmark results for you all.
The results are from test runs between the machines thyra (Aarhus,
Denmark), bazooka (LA, USA), and serengeti (Trondheim, Norway). I have
used benchmark.py, some shell scripting and lots of patience :-)
inline: mul-par.png
The first graph shows parallel multiplications in the standard 65 bit
field. The time is the total time used on the three machines. One can
see that bazooka is a bit slower than the two others, but overall they
follow each other nicely.
inline: mul-par-normalized.png
The second graph gives the normalized results from the first graph, that
is, the time per multiplication. The time stabilized around 1.5 ms per
multiplication when doing more than 2000 multiplications.
inline: mul-seq.png
If we run things sequential instead, we get the results in graph 3. Here
I have only tested up to 1000 multiplications with 10 repetitions since
that took 30 minutes.
inline: mul-seq-normalized.png
Again, we can look at the time per multiplication. I have not good
intuition as to why the two fast machines (thyra and serengeti) are able
to do the first 10 multiplications much faster than the later ones. The
bazooka machine is more stable.
I believe these are the most comprehensive results I have made with VIFF
so far. For the parallel multiplications I repeated the test at least
100 times and used the median value for the graphs. I took the median
value instead of the average value since the median should be more
robust when the data has outliers. But since I have no clue about
statistics, I would love to hear other suggestions.
--
Martin Geisler
___
viff-devel mailing list (http://viff.dk/)
viff-devel@viff.dk
http://lists.viff.dk/listinfo.cgi/viff-devel-viff.dk
