/home/mmadon/demand-response-user/scripts/util.py:93: UserWarning: Creating legend with loc="best" can be slow with large amounts of data.
fig.savefig(expe_file + '_viz.png')
/nix/store/8nqifk7a8fq7cb4i9rhhs86bn7aqmmrs-python3.8-ipython-7.21.0/lib/python3.8/site-packages/IPython/core/pylabtools.py:132: UserWarning: Creating legend with loc="best" can be slow with large amounts of data.
We open all the experiment outputs, compute raw metrics such as energy in window or slowdown / waiting time, and save this as a csv file. Also compute these metrics for each experiments and each user behavior relatively to the baseline behavior.
Descriptive stats: for each expe and each window, calculate the number of jobs in dr_window et infrastructure load in dr_window. Use the baseline expe to get these stats.
plt.suptitle("(a) number of jobs in window; (b) total mass (in core-hour) in window; (c) total mass in window by number of jobs (1-hour window); (d) total mass in window by weekday (1-hour window)")
plt.show()
```
%% Output
%% Cell type:markdown id:ed775f60 tags:
Analysis:
- no correletion between the number of jobs in the demand response window and the average load during that period: the jobs have very different size
- data very scattered, even if we notice that in more than 50% of the case, the infrastructure is at almost full load during the demand response event, and at more than 80% load in more than 75% of the case
- the 1-hour window (from 16:00 to 17:00) is more loaded on average than the 4-hour window (from 16:00 to 20:00)
- when the platform is fully loaded (right-end side of graphs): involving the user does not help, because the queue of jobs needs to empty itself: effets de latence et d'inertie
- no obvious correlation between the load of the platform during the DR event and the energy gain relative to the baseline (confirmed by the graph below)
- une prise en compte niveau scheduler de l’effort utilisateur indispensable