| |
Marmousi 2D
MARMOUSI - 2D
(2nd order time, 8th order space)
 |
 |
| Velocity Model |
Migrated Image |
239 shots at:
39.0 seconds / shot (single GT200)
FUTURE seconds / shot (x CPU cores, TBD GHz)
Document Revision: V1.30
Last Updated: December 10, 2008
RTM Orders: 2nd Time, 8th Space
SW Release Used: “Columbia”
SW Release Date: November 7, 2008
Benchmark Status: Public, Non?Confidential
RTM Model Definition
| Problem Definition |
Parameters |
|
|
| 2D line length |
9.2 |
km |
|
| Survey depth |
3.0 |
km |
|
| Total number of shots |
239 |
shots |
|
| Number of receivers, per shot |
96 |
receivers/shot |
|
| |
|
|
|
| Minimum shot offset |
200.0 |
m |
|
| Maximum shot offset |
2,575.0 |
m |
|
| |
|
|
|
| Trace sampling interval |
4.0 |
milliseconds |
|
| Trace length |
3.0 |
seconds |
|
| Number of samples |
750 |
samples |
|
| |
|
|
|
| Migration frequency |
50.0 |
Hz |
|
| Maximum velocity |
5,500 |
m/s |
|
| Minimum velocity (water) |
1,500 |
m/s |
|
| Maximum wavelength |
110 |
m |
[fn for migration freq and max. velocity] |
| Minimum wavelength |
30 |
m |
[fn for migration freq and min. velocity] |
| |
|
|
|
| Temporal order |
2 |
nd |
|
| Spatial order |
8 |
th |
|
| Grid points per wavelength |
4.0 |
points/lambda |
[dispersion-free for minimum wavelength] |
| Maximum allowable cell size (for dispersion) |
7.5 |
m |
[dispersion-free for minimum wavelength] |
| Working cell size |
4 |
m |
[cell size actually used in the migration] |
| |
|
|
|
| dt |
0.3 |
miliseconds |
|
| Simulation timesteps |
10,000 |
timesteps |
|
| |
|
|
|
| RTM Model Assumptions |
Parameters |
|
|
| Maximum volumes in memory |
6 |
|
[6, +1 for density, +6 for (VTI) anisotrophy] |
| |
|
|
|
| RTM Boundaries |
|
|
|
| Number of sponge boundary layers, each x-axis |
30 |
|
|
| Number of sponge boundary layers, Zmin |
30 |
|
|
| Number of sponge boundary layers,Zmax |
30 |
|
|
| Extra cells at EACH boundary due to FD stencil |
4 |
|
[2 cells for 4th-order-in-space, 4 cells for 8th-order-in-space] |
| |
|
|
|
| Size of RTM domain |
1.73 |
million cells |
|
| Size of RTM domain, including boundaries |
1.94 |
million cells |
|
| Memory high watermark, including boundaries |
44.3 |
Mbytes |
[given num. volumes stored, single precision] |
| |
|
|
|
| Calculations |
|
|
|
| "Normal" cells, X |
2,300 |
|
|
| "Normal" cells, Z |
750 |
|
|
| Total "normal" cell count |
1,725,000 |
|
|
| Cells including boundaries, X |
2,368 |
|
|
| Cells including boundaries, Z |
818 |
|
|
| Total Cell Count, including boundaries |
1,937,024 |
|
|
| Memory occupation, including boundaries |
46,488,576 |
bytes |
|
| Additional fields stored |
424,048 |
|
|
| Memory occupation, additional storage required (MB) |
1.62 |
Mbytes |
|
GPU Performance
| RTM Parameters (carry over) |
|
|
|
| Size of RTM domain, including boundaries |
1.94 |
million cells |
|
| Memory high water mark, including boundaries |
44.3 |
Mbytes |
|
| Total number of shots |
239 |
shots |
|
| |
|
|
|
| Single GPU specifications |
|
|
|
| GPU memory (GB) |
3.6 |
GB |
[next gen h/w from NVIDIA, GT200 = 4GB with 0.9 Safety Factor] |
| |
|
|
|
| "Accelerated Node" definition |
|
|
|
| CPU Server hosts per "Tesla" (GPU Server) |
2 |
|
NOT USED |
| Number of GPU servers |
1 |
|
NOT USED |
| Individual GPU's per GPU server |
4 |
|
NOT USED |
| |
|
|
|
| Single GPU Configuration |
|
|
|
| Total time to migrate single shot |
39.0 |
seconds |
[Single NVIDIA GT200] |
| Time to migrate all shots |
2.6 |
hours |
[Assumes none of the shots are migrated on parallel equipment] |
| |
|
|
|
| Forward Propagation (and Illumination Calculation) |
|
|
|
| Total time for forward pass |
14.2679 |
seconds |
|
| Time per iteration (forward) |
0.00142679 |
seconds |
[Acceleware recommends that this is the best metric to use for performance comparisons] |
| Throughput (NOT including boundary cells) |
1,209 |
Mcells/s |
[Boundary cells are included in the time but not in the number of cells updated] |
| Throughput (INCLUDING boundary cells) |
1,358 |
Mcells/s |
|
| |
|
|
|
| Reverse Propagation (inc. Correlation) |
|
|
|
| Total time, for reverse pass |
24.7583 |
seconds |
|
| Time per iteration (reverse) |
0.00247583 |
seconds |
[Acceleware recommends that this is the best metric to use for performance comparisons] |
| Throughput (NOT Including Boundary Cells) |
697 |
Mcells/s |
[NB Two identically sized RTM domains exist in memory, at the same time, during this stage of the migration. Cells for only one of the domains are counted] |
| Throughput (INCLUDING Boundary Cells) |
782 |
Mcells/s |
|
Summary of Benchmark Results
| |
|
|
Size of Domain |
|
Performance |
| |
|
|
|
|
|
"Normal" Cells |
Including Boundaries |
|
|
|
Time per iteration |
| |
Temporal Order |
Spacial Order |
X |
Y |
Z |
(Mcells) |
(Mcells) |
Memory High Water Mark |
Number of GPU's per shot |
Total Time, per Shot |
Forward
(seconds) |
Reverse
(seconds) |
| BP Dataset 2D |
2nd |
8th |
16,850 |
- |
2,975 |
50.1 |
51.5 |
1.15 Gbytes |
1 |
Under Construction |
| Foothills Dataset - 2D |
2nd |
8th |
6,250 |
- |
2,500 |
15.6 |
16.2 |
371.4 Mbytes |
1 |
35.8 seconds |
0.0013035 |
0.0022724 |
| Marmousi 2D |
2nd |
8th |
2,300 |
- |
750 |
1.73 |
1.94 |
44.3 Mbytes |
1 |
39.0 seconds |
0.0014268 |
0.0024758 |
| SEG Salt, A1 - 3D |
2nd |
4th |
100 |
676 |
185 |
12.5 |
30.2 |
691.6 Mbytes |
1 |
144.8 seconds |
0.0222250 |
0.0357102 |
| SEG Salt, A2 - 3D |
2nd |
4th |
676 |
50 |
185 |
6.3 |
21.0 |
480.8 Mbytes |
1 |
116.7 seconds |
0.0175312 |
0.0291397 |
| SEG Salt, B1 - 3D |
|
|
Under Construction |
|
Under Construction |
Click here for more information on RTM Benchmarks or email us at seismic@acceleware.com
|
 |
