We now give some CM timing results of our Kirchhoff 3-D prestack time migration algorithm.
CM TIMING: 1 3D shot gather --> 64x64x256 migration voxels
518 oas> cake cmig cmattach -w -S0-1 /r1/david/bin/SUN4/cmig.x Attached to Connection Machine System MAUNALOA cold booting ... done. Attached to 8192 processors on sequencers 0-1, microcode version 6002 Paris safety is off. CM clock speed: 7.00 MHz
Migration complete Timings from timer # 0 Total Mflops =279345.9 in elapsed time= 654.662178 and CM time= 654.106250 Total elapsed Mflps/s: 8192 proc= 426.702349 -> 64k proc= 3413.618792 Total CM Mflps/s: 8192 proc= 427.065006 -> 64k proc= 3416.520047 MIN migration value -2787.723 MAX migration value 5700.036 FORTRAN STOP Detaching... done. 620.4u 32.1s 11:38 93% 0+1832k 615+16io 720pf+0w
The above timing is for a single 3-D prestack shot migration described in the point diffractor example. The migration was run using both sequencers for a full 8192 processors with 64 Mb memory. The shot migration required 279 Gflop total, which was performed by the CM in 654 sec (10.9 min), for a total CM computational rate of 427 Mflop/s. The flop count was measured as 65 flops per image voxel per trace, which included sqrt operations as 8.4 flop each (Biondi, 1991, p. 69-70). This timing scales to 3.4 Gflop/s on a full 64k processor CM, provided that eight shot gathers (or 8 times as many total trace samples) are available to saturate CM core memory, which is obviously no problem at all with 3-D seismic data! This timing includes all migration calculations, cshifting, internal memory data access and processor-to-processor data movement. However, it does not include external I/O to and from the front-end, which tends to be variable depending on unrelated front-end loads, and would change dramatically if one used CM's data vault or a high-speed disk array. For this 11 cpu minute shot migration, typical elapsed times including front-end I/O average about an extra 20-30%, or 13-14 minutes total elapsed time per 3-D prestack shot migration.