High Frequency CyberShake - Revision history

Scottcal at 21:34, 2 December 2014

2014-12-02T21:34:03Z

2014-11-22T00:15:44Z

2014-11-22T00:14:44Z

2014-11-22T00:13:14Z

Rupture Generation code (genslip v3.3.1)

2014-11-22T00:05:50Z

2014-11-12T21:37:39Z

Random vs uniform hypocenter distribution

2014-11-12T21:37:24Z

Random vs uniform hypocenter distribution

2014-11-12T20:12:51Z

2014-11-12T20:08:22Z

2014-11-12T19:36:40Z

Rupture Generation code (genslip v3.3.1)

@@ Line 69: / Line 69: @@
 * TODO: Replace get_master_list with get_master_list_opt from 1 Hz extract SGT code
 * TODO: Replace find_sgt linear search with binary search
-* TODO: Replace time-domain convolution with FFT/frequency convolution/IFFT.
+* Tried replacing time-domain convolution with frequency convolution using FFTW.  Ended up being 33% slower.  I kept the capability in as a flag, as it may be useful later since we could calculate the FFT once, then perform multiple convolutions for different rupture variations.
 = RotD calculations =

@@ Line 37: / Line 37: @@
 == Random vs uniform hypocenter distribution ==
-[[File:Comparison of hypocenter distributions 3sec.png|thumb|400px|Figure 3<br\>Black is random, purple is uniform.]]
+[[File:Comparison of hypocenter distributions 3sec.png|thumb|400px|Figure 3<br\>3 sec SA for WNGC. Black is random, purple is uniform.]]
-[[File:Comparison of hypocenter distributions 5sec.png|thumb|400px|Figure 4<br\>Black is random, purple is uniform.]]
+[[File:Comparison of hypocenter distributions 5sec.png|thumb|400px|Figure 4<br\>5 sec SA for WNGC. Black is random, purple is uniform.]]
-[[File:Comparison of hypocenter distributions 10sec.png|thumb|400px|Figure 5<br\>Black is random, purple is uniform.]]
+[[File:Comparison of hypocenter distributions 10sec.png|thumb|400px|Figure 5<br\>10 sec SA for WNGC.  Black is random, purple is uniform.]]
 We conducted an experiment to see what the impact of random vs uniform hypocenter distribution is on hazard curves.  Plots comparing run 3847 (random) to run 3837 (uniform) are in Figures 3-5.

@@ Line 51: / Line 51: @@
 * Replaced bubble sort of sort_rake with qsort call:  0m25.547s (92.3% improvement)
 * Replaced fourg FFT with FFTW FFT, copied data over to FFTW structures: 0m23.045s (9.8% incremental improvement, 93.4% cumulative)
-* Removed sort entirely, since median is no longer used: 0m23.140s (0.4% worse)
+* Removed sort entirely, since median is no longer used: 0m23.140s (0.4% worse, probably noise)
 = Seismogram synthesis code =

@@ Line 16: / Line 16: @@
 Cyan = Run_ID 2507: 200m SGTs, 1000m synth (most points ignored)]]
-To move to higher frequencies, the rupture geometries must have a denser grid spacing, from 1000m for 0.5 Hz to 200 m for 1.0 Hz.  We adopted genslip v3.3.1 to resolve an issue in which, since the dimensions of the rupture surfaces can vary a bit going from 1000m to 200m, some ruptures went from 8 rupture variations to 2, resulting in undersampling the rupture and causing artificially large ground motions (Figure 1; magenta curve, compared to green and cyan):
+To move to higher frequencies, the rupture geometries must have a denser grid spacing, from 1000m for 0.5 Hz to 200 m for 1.0 Hz.  We adopted genslip v3.3.1 to resolve an issue in which, since the dimensions of the rupture surfaces can vary a bit going from 1000m to 200m, some ruptures went from 8 rupture variations to 2, resulting in undersampling the rupture and causing artificially large ground motions (Figure 1; magenta curve, compared to green and cyan).  Figure 1 illustrates that the issues is related to the resolution of the fault surfaces used in the synthesis, not the resolution the SGTs are saved at.
 To address this issue, we moved to genslip v3.3.1, which has the following modifications:

@@ Line 56: / Line 56: @@
 == Batching for 1 Hz ==
 We took the batching capability from SeisPSA_multi and integrated it into the large memory version for 1 Hz calculations, producing SeisPSA_lm_multi.  In this way we can batch our 1 Hz post-processing, reducing the input I/O by a factor equal to the batching factor.
 We verified this code by comparing run 3845 (PAS, 0.5 Hz, ERF 35) to run 2722 (PAS, 0.5 Hz, ERF 35) (Figure 6).
 == Optimizations ==
@@ Line 71: / Line 72: @@
 = RotD calculations =
 As the UGMS would like to see RTGM results which use RotD100, we integrated the RotD code into SeisPSA_lm_multi as an optional flag.  We verified this by comparing the RotD100 curves from run 3845 (PAS, 0.5 Hz, ERF 35, RotD) to run 2722 (PAS, 0.5 Hz, no RotD) (Figure 7).

@@ Line 60: / Line 60: @@
 We verified this code by comparing run 3845 (PAS, 0.5 Hz, ERF 35) to run 2722 (PAS, 0.5 Hz, ERF 35) (Figure 6).
-[[File:Verification of SeisPSA_lm_multi.png|thumb|left|400px|Figure 6]]
+[[File:Verification of SeisPSA_lm_multi 3sec.png|thumb|left|400px|Figure 6]]
 == Optimizations ==
@@ Line 72: / Line 72: @@
 = RotD calculations =
-As the UGMS would like to see RTGM results which use RotD100, we integrated the RotD code into SeisPSA_lm_multi as an optional flag.  We verified this by comparing the RotD100 curves from run 3845 (PAS, 0.5 Hz, ERF 35, RotD) to run 2722 (PAS, 0.5 Hz, no RotD) (Figure 5).
+As the UGMS would like to see RTGM results which use RotD100, we integrated the RotD code into SeisPSA_lm_multi as an optional flag.  We verified this by comparing the RotD100 curves from run 3845 (PAS, 0.5 Hz, ERF 35, RotD) to run 2722 (PAS, 0.5 Hz, no RotD) (Figure 7).
-[[File:Verification of RotD code.png|thumb|left|400px|Figure 5]]
+[[File:Verification of RotD code 3sec.png|thumb|left|400px|Figure 7]]

@@ Line 36: / Line 36: @@
 == Random vs uniform hypocenter distribution ==
 [[File:Comparison of hypocenter distributions 3sec.png|thumb|400px|Figure 3]]
@@ Line 44: / Line 42: @@
 [[File:Comparison of hypocenter distributions 10sec.png|thumb|400px|Figure 5]]
 == Optimizations ==