About the run time in cloud computing scenario my measure might be useless. A cloud instance is a VM that runs on I don't know what hardware and might be shared with I don't know who. So even if inside my VM the CPU was idle, I don't now how the CPU was used by other VMs. That being said, for our particular cloud service provider we observe (on a daily basis for several months) that the performances are below what we have on our modern workstations.

Just to be clear : I'm neither complaining about graph tool performances, nor thinking about improvements that would be cloud computing driven. I just wanted to make clear that modest improvement (as measured on a modern work station) might be very meaningful in another environment.

In order to profile the initialization I moved the initialization code in two functions init_pmap and init_dist, and did 100 runs of shortest_distance with identical parameters, cProfiles gives :

23484 function calls (23331 primitive calls) in 6.733 seconds

Ordered by: cumulative time

ncalls tottime percall cumtime percall filename:lineno(function)

100 3.723 0.037 6.731 0.067 graph_tool/topology/__init__.py:1512(shortest_distance)

100 0.001 0.000 1.833 0.018 graph_tool/topology/__init__.py:1507(init_pmap)

100 0.000 0.000 1.833 0.018 graph_tool/decorators.py:1(copy_property)

200/100 0.003 0.000 1.832 0.018 graph_tool/decorators.py:126(wrapper)

100 1.275 0.013 1.827 0.018 graph_tool/__init__.py:2348(copy_property)

100 0.001 0.000 1.163 0.012 graph_tool/topology/__init__.py:1500(init_dist)

100 0.000 0.000 0.645 0.006 graph_tool/__init__.py:661(<lambda>)

100 0.643 0.006 0.644 0.006 graph_tool/__init__.py:614(__get_set_f_array)

700 0.001 0.000 0.540 0.001 graph_tool/__init__.py:166(_prop)

700 0.003 0.000 0.538 0.001 graph_tool/__init__.py:392(_get_any)

700 0.534 0.001 0.534 0.001 graph_tool/__init__.py:814(reserve)

200 0.000 0.000 0.518 0.003 graph_tool/__init__.py:559(get_array)

200 0.515 0.003 0.517 0.003 graph_tool/__init__.py:581(_get_data)

200 0.004 0.000 0.011 0.000 graph_tool/__init__.py:2289(new_vertex_property)

200 0.002 0.000 0.010 0.000 graph_tool/__init__.py:3071(__init__)

600 0.002 0.000 0.008 0.000 graph_tool/__init__.py:377(__init__)

200 0.002 0.000 0.007 0.000 graph_tool/__init__.py:1531(__init__)

100 0.000 0.000 0.006 0.000 graph_tool/__init__.py:2274(new_property)

800 0.002 0.000 0.004 0.000 graph_tool/__init__.py:202(_type_alias)

599 0.001 0.000 0.003 0.000 graph_tool/__init__.py:437(__del__)

600 0.001 0.000 0.002 0.000 graph_tool/__init__.py:264(_converter)

599 0.002 0.000 0.002 0.000 graph_tool/__init__.py:432(__unregister_map)

200 0.001 0.000 0.002 0.000 graph_tool/__init__.py:909(__new__)

The complete cprofile output is attached to this mail.

Le jeu. 29 juin 2017 à 13:16, Tiago de Paula Peixoto <tiago@skewed.de> a écrit :

On 29.06.2017 11:58, François Kawala wrote:
> I did the above timings on a modern work station, to switch to a cloud
> provider induces a non negligible overhead (+63% for the distances / +469%
> for the pred map).

This I don't understand at all. Why would the behavior in a "cloud"
environment be any different?

--
Tiago de Paula Peixoto <tiago@skewed.de>

_______________________________________________
graph-tool mailing list
graph-tool@skewed.de
https://lists.skewed.de/mailman/listinfo/graph-tool