Node Level Scaling on Haswell
On this page, we track the state of node-level scaling on a Haswell machine.
The following graphs have been created on a 20 core Haswell system from running Octotiger for 7 levels of refinement with this command-line:
octotiger -Disableoutput -Stopstep=1 -Problem=moving_star -Xscale=32 -Odt=0.5 -Maxlevel=7
The measurements represent the snapshot taken at different dates in the past.
Please note, that starting with 2/12/2017 the shown overall times do not include the final output traversal anymore. This conveys the wrong impression of a massive improvement in runtime and efficiency in the time frame between 2/5/2017 and 2/12/2017 for the first two plots.
The shown data for the time step execution time and efficiency are directly comparable, however.
The Overall execution time:
The parallel efficiency calculated from the overall execution time:
The Average Time for a timestep:
The parallel efficiency calculated from the average execution time of the timesteps:
