OpenCL
No news.
GPU Servers
CI Server:
- Setting PCIe version in BIOS fixed Nvidia GPU
- Handed node to Giulio and Sergio for CI integration
- Will ping Sergio if he has an update
Highly Ionizing Particles
Ernst fixed regression in FST when running simulation.
Opened PR for the current implementation.
Filter Performance
Checked filter performance on 50khz pbpb timeframe. Matched zeroed values to MC labels.
Good news: Filter appears to work.
Bad news: Doesn't seem to work very well.
Confusion matrix
| Cutoff Threshold |
Correctly Zeroed |
Tails Leaked |
Correctly Kept |
False Positive |
Precision (# correctly zeroed / # total zeroed) |
| 150 ADC |
0.3910 |
0.6090 |
0.9977 |
0.0023 |
0.5370 |
| 100 ADC |
0.3959 |
0.6041 |
0.9977 |
0.0023 |
0.5362 |
| 50 ADC |
0.4115 |
0.5885 |
0.9974 |
0.0026 |
0.5161 |
Some example (original data), Yellow regions mark saturated ADC
Data overlayed with MC (flood fill around saturated ADC)
Data with filter applied
Runtime Overhead
- Noisy Pad Filter runs only on first fragment by default -> makes no sense when HIP filter is enabled
- Total overhead when running with filter on all fragments: 10% walltime
Other Points
- Ideas to improve filter performance:
- Start some timebins backwards from trigger to handle rising edge
- Moving average / exponential filter for threshold to better handle noise in tail
