While using alumet for measuring energy consumption, we recently came about understanding the importance of metric Types.
I believe you'll have carefully considered how these metric Types influence computations involving interpolation within metrics, convolution of two metrics, etc. However, when we tried to compute the total energy of a process (CPU + GPU), we came across the following two points:
1. Documentation Bug:
In https://alumet-dev.github.io/user-book/plugins/2_transforms/energy-attribution.html, the Type of metric should be Counter Diff. Because it is convolved using two metrics, cpu_time_delta and rapl_consumed_energy, both of which are Counter Diff. IIUC convolving a Counter Diff with another Counter Diff/Gauge should result in a Counter Diff
2. Feature Request
It would be nice to have the Type of the metric also stored in the Outputs (CSV, Prometheus), this allows easier downstream analysis.
Related visuals to understand are here: thealanjason/alumet-insight#30
While using alumet for measuring energy consumption, we recently came about understanding the importance of metric
Types.I believe you'll have carefully considered how these metric
Types influence computations involving interpolation within metrics, convolution of two metrics, etc. However, when we tried to compute the total energy of a process (CPU + GPU), we came across the following two points:1. Documentation Bug:
In https://alumet-dev.github.io/user-book/plugins/2_transforms/energy-attribution.html, the Type of metric should be
Counter Diff. Because it is convolved using two metrics, cpu_time_delta and rapl_consumed_energy, both of which areCounter Diff. IIUC convolving aCounter Diffwith anotherCounter Diff/Gaugeshould result in aCounter Diff2. Feature Request
It would be nice to have the
Typeof the metric also stored in the Outputs (CSV, Prometheus), this allows easier downstream analysis.Related visuals to understand are here: thealanjason/alumet-insight#30