Issue
List of commands for this test
Code for time monitoring
Major histograms
Results
t point | time increment | point description | time for rank 0/1 | rank 0/80 | rank 30/80 | rank 60/80 |
---|---|---|---|---|---|---|
1 | t1 - t0 | det.raw | 0.8±0.2 ms | 4.0 ±0.6 ms | 3.2±0.4 ms | 3.5 ±0.8 ms |
2 | t2 - t1 | det.pedestals | 15±3 μs | 36 ±10 μs | 31±6 μs | 39 ±17 μs |
3 | t3 - t2 | det.gain,offset | 15±2 μs | 27 ±4 μs | 26±4 μs | 27 ±6 μs |
4 | ... | cmpars | 25±1 μs | 50 ±7 μs | 58±26 μs | 71 ±33 μs |
5 | gfac | 2±0 μs | 6 ±1 μs | 7±1 μs | 7 ±2 μs | |
6 | gr0,1,2 | 1.3±0.2 ms | 10.5 ±1.1 ms | 7.0±0.9 ms | 9.7 ±1.6 ms | |
7 | make arrf | 1.76±0.05 ms | 9.2 ±0.9 ms | 6.3±0.7 ms | 9.0 ±1.5 ms | |
8 | subtract peds | 93.7±3.1 ms | 191 ±11 ms | 181±15 ms | 259 ±26 ms | |
9 | eval gain factor for gain ranges | 4.9±0.6 ms | 20.3 ±1.5 ms | 14.6±1.2 ms | 17.3 ±2.0 ms | |
10 | eval offset for gain ranges | 6.2±0.4 ms | 18.5 ±1.3 ms | 18.4±1.4 ms | 19.2 ±2.1 ms | |
11 | subtract offset | 1.0±0.2 ms | 6.0 ±0.7 ms | 5.3±0.6 ms | 6.2 ±1.2 ms | |
12 | get mask | 3±2 μs | 6 ±2 μs | 6±2 μs | 7 ±2 μs | |
13 | common mode turned off | 7±1 μs | 15 ±2 μs | 17±2 μs | 20 ±3 μs | |
14 | t14 - t13 | apply gain factor and mask | 4.0±0.7 ms | 14.9 ±2.0 ms | 13.9±1.6 ms | 19.2 ±3.5 ms |
99 | t14 - t0 | per evt time, inside det.calib | 109.8±4.2 ms | 276 ±15 ms | 247±13 ms | 345 ±29 ms |
0 | t0 - t0 previous evt | time between consecutive det.calib | 115.4±3.9 ms | 335 ±16 ms | 307±14 ms | 398 ±32 ms |
Summary
- single core processing is faster than per/core time in 80–core case, factor 2.5-3 FOR ALL OPERATIONS
- in 80-core case: time per core is consistent between cores
- all constants are cashed and access to constants is fast at sub-milisecond level
- common mode correction is turned off, as well as mask?
- most time consuming operation is indexed pedestal subtraction
indexed by gain ranges pedestal subtraction
t07 = time() arrf[gr0] -= peds[0,gr0] arrf[gr1] -= peds[1,gr1] arrf[gr2] -= peds[2,gr2] t08 = time()
- bad single-to-multicore scaling issue has nothing to do with particular algorithm, it is common problem for any algorithm
References
Overview
Content Tools