More suggestions for PDF report
Hi to everyone,
yesterday I discussed with Thomas the PDF report and we have some further suggestions:
Job overview
- The nodelist should be placed on the left hand side
- The spaces between for example "User name" and ":" should be deleted
- The used wallclock time in the Job overview and in the "Global summary" should be identical (both should have the same number of decimal places)
- "Time of job submition" --> "Time of job submission"
Global summary of resource usage
- At the moment all bars are green. Is there a traffic light system implemented with red - yellow - green?
- Why swap sum? It is possible, that Linux swaps part of its kernel datastrucures to disk, because it's not needed. In that case we have swap activities, but the reason is not our job. Why not swap per time unit?
- Memory and Swap in percent? (additionally)
- GPU memory in GB
- Memory in GiB or GB (analogously MB, etc)?
- Network and I/O with seperate bars for sent and receive?
- In general it would be interesting to have an bar with the whole interval a metric can occupy, and to draw those regions with different colours dependent if the value is good, bad, etc. And finally we can mark the value the job has in this metric with an arrow, which points to that place in the bar, where our value is placed (see diagramm in the ISC poster 2017).
- GPU values to the other values and not in 2 separarted blocks
Recommendations
- New page (2nd page)
Node Distribution
- Swap Avg and stddev column?
- I wonder, if mean + stddev < max (or > min) must always hold. Could it be, that the stddev was divided by n-1 and not by n?
- In the box plots: what is the orange line? The mean or average value? Perhaps a short description of the boxplots at the boxplots or in its vicinity?
- The length of the boxplots is different
- Perhaps we should think about the problem, that the boxplot is not very meaningful in the case of 3 or fewer nodes. In that case a bar chart instead of a boxplot would be advantageous?
Node timeseries plot
- If there are many nodes in the job a visualization problem will occur. How can we handle that. Visualize only a sample of the nodes or grouping the nodes?
- Node CPU uage: the Interval should be from 0 to 100 % and should be cut.
GPU distribution summary:
- Why power limit? It's a constant value!?
- Memory RSS HWM: Device or host mem?
- CPU usage: 856.00 % (mean value). Different to CPU usage in the first box plot table!
GPU timeseries
- Report of JobID 2365880 (GPU Job): The first 3 rows are a bit strange, especially GPU usage vs GPU processes