`(EXCERPTED)
`
`
`
`
`
`
`
`
`
`Elernenrsflf
`3-I] Seisrnulngu
`
`Ex. PGS 1036
`
`
`
`Elements of
`3-H Seismulugu
`
`is
`
`.
`
`.
`
`.1 .-A;v:%.
`
`EX. PGS 1 036
`
`Ex. PGS 1036
`
`
`
`Copyright © 1999 by Christopher L. Liner
`All inquiries should be referred to PennWel1 Publishing
`1421 South Sheridan/P.O. Box 1260
`Tulsa, Oklahoma 74101
`
`
`
`Cover Design by Matt Berkenbile
`Book Layout by Geoff Harwood with Stormgrafx
`
`All rights reserved. No part of this book may be reproduced, stored in a
`retrieval system, or transcribed in any form or by any means, electronic or
`mechanical, including photocopying and recording, without the prior written
`permission of the publisher.
`
`Printed in the United States of America
`
`12345040302010O
`
`iv
`
`EX. PGS 1 036
`
`Ex. PGS 1036
`
`
`
`
`
`§ 1 E
`
`
`
`__._.4—.£-(x4:-~.Amu-.—.——-—-——-j<-—-——-——---
`
`<1?
`
`IElements of3-D Seismology
`
`CD
`
`GMP Fold
`
`For a 3-D survey to yield good data quality, the target fold should
`be about one-half of the fold required to shoot good 2-D data in the
`area. This is a result of migration and dip moveout which result in more
`mixing of 3-D data than occurs in 2-D.
`Some points on fold
`
`1. High fold costs more at acquisition time
`2. Low fold (< 10) 3-D has been successful
`3. Lower fold with right bin size may be better than high fold
`with too large a bin
`
`Spatial aliasing is an effect of trace spacing relative to frequency,
`velocity, and slope of a seismic event. With adequate trace spacing, the
`points along a seismic event are seen and processed as part of the con-
`tinuous event. When trace spacing is too coarse, individual points do
`not seem to coalesce to a continuous event, which confuses not only the
`eye but processing programs as well. This can seriously degrade data
`quality and the ability to create a usable image.
`Figure 7-6 shows one way of defining spatial aliasing. In this view
`spatial is based on trace-to-trace delay associated with a clipping reflec-
`tor. Since the delay is related to trace spacing, the issue is really one of
`midpoint interval. This, in turn, is related to shot and receiver interval.
`For 2-D data, mid oint s acin , A4,-, shot interval, Si, and receiver
`P
`P
`g
`group interval, R,-, are related by
`
`
`
`M, = % Mz'n(S,-, R,-)
`
`To avoid spatial aliasing on the stack section we require
`
`M, <
`
`4 fm Sz'n0
`
`
`
`(7.10)
`
`(7.11)
`
`EX. PGS 1 036
`
`Ex. PGS 1036
`
`
`
`
`
`Chapter 7 Survey Predesignj(I)
`
`where um, is the interval velocity near (or immediately above) the target,
`fmx is maximum signal frequency and 0 is the physical dip angle of the
`reflecting bed. When interval velocity is not known, average velocity can
`be used. But this will give a bin size estimate that is smaller than
`required. The design condition is rightly based on maximum frequen-
`cy, but the dominant frequency is often used. This means high fre-
`quency components risk being aliased.
`Spatial aliasing is not difficult to recognize on real data, (Figure 7-
`7). The main problem with spatial aliasing is the detrimental effect it
`has on two very expensive processes: dip moveout and migration.
`Figures 7-8 and 7-9 give a migration example.
`We note for design purposes that diffraction limbs (Figure 7-4)
`appear as 0 = 90° events. The Sin6 term is sometimes invoked to justify
`a non—square bin in 3-D shooting. However, this increases risk of spa-
`tially aliasing the data, so the safe design rule is to use 6 = 90°. In this
`case, the midpoint spacing condition reduces to M < /\,,,,,,,/4. This agrees
`with the minimum bin size requirement, Adm/4, for a 3-D survey as dis-
`
`cussed in chapter 14.
`The unaliased midpoint interval grows with depth due to increas-
`
`ing velocity and decreasing frequency.
`Some points on spatial aliasing
`
`1. Safe direction is smaller
`
`2. Will be different for shallow and deep targets
`3. Use fma, and 6 = 90° for safest midpoint interval
`4. Smaller midpoint interval costs more at acquisition time
`5. Same equation for 2-D midpoint interval and 3-D bin size
`
`
`
`__.....__,..;5.n_=....4-45.4_
`
`
`
`EX. PGS 1 036
`
`Ex. PGS 1036
`
`
`
`<13
`
`JElements of3-D Seismology
`
`(I)
`
`
`
`
`
`2-ll Suruev size
`
`This calculation shows how much disk space is needed to store a
`moderate 2-D seismic line in SEGY format:
`
`lsample value
`6 secs data @ .004 sec sample rate
`1,500 * 4 + 240(header)
`96 channel * 800 shots
`76,800 traces * 6,240" bytes/trace
`add line headers 3600 bytes
`479,232,000 / (1,024)"2
`
`=
`=
`=
`=
`=
`~
`
`4 bytes (floating point \#)
`1,500 samples/trace
`6,240 bytes/trace
`76,800 traces
`479,232,000 bytes
`479,235,600 bytes
`457 Mb
`
`From the typical RAM and disk storage numbers above, we con-
`clude that the 2-D seismic line can fit
`
`1. in RAM of a large wk or small sc
`
`2. on disk of a wk or pc
`
`The RAM issue is important because some processes (e.g.,
`prestack migration) want to hold the entire data set, plus work space, in
`RAM. If it can't all be stuffed into RAM, we are left with a significant
`
`data management problem.
`
`3-n Suruev Size
`
`Do things get better for 3-D seismic data? Hardly. Large 3-D sur-
`veys today can contain hundreds of millions of prestack traces. This cal-
`culation shows the size of a medium—sized 3-D data set composed of 500
`
`2-D lines like the one above:
`
`.
`
`.
`
`206
`
`EX. PGS 1 036
`
`we
`
`._,|1i,
`
`Ex. PGS 1036
`
`
`
`Chapter 77 Computing NeedsJ(1)
`
`[38,400,000 traces!]
`
`500 2-D lines
`
`500 lines * 479,235,600 bytes/line
`2.39616*10"11 bytes
`228 515 Mb
`223 Gb
`
`12|||l
`
` (1)
`
`This amount of data will not fit in RAM of any existing sc (not
`even close), but we could probably find enough pasture for it on the
`disk farm.
`
`P|'0GESSiIl9 Sllflflll
`
`RAM and disk storage are part of the seismic computing bottle-
`neck; the other problem is processing speed. It is of little use to squeeze
`a mass of seismic data into computer RAM if, once there, it takes 2 or 3
`years to process it.
`The fundamental unit of processing is the flop. To confuse the
`uninitiated, the flop is both an operation, floating point operation, and a
`rate, floating point operations per second. Here is a table of processing
`speed units.
`
`t
`
`flop
`
`=
`=
`=
`sqrt()
`1 megaflop =
`1 gigaflop
`=
`1teraflop
`=
`1 petaflop
`=
`1 exaflop
`=
`
`floating point operation + */
`floating point operation per second
`10 15 floating point operations
`10"6 flop =1 Mflop
`10"9 flop = 1 Gflop
`10"12 flop =1Tf|op
`10/‘15f1op =1 Pflop
`10"18 flop =1 Eflop
`
`As with petabyte and exabyte, the last two terms are not yet in
`common use. We don't have machines that fast. The current world
`speed record for computer processing is around 1 teraflop in perform-
`ance.
`
`The alert reader will see inconsistent use of prefixes between
`megabyte and megaflop. A megabyte is 1,0242 bytes, but a megaflop is
`1,0002 flops.
`There are many different approaches to achieving high flop rates.
`None of them are cheap. Some typical, and not so typical, processing
`speeds are:
`
`
`
`EX. PGS 1036
`
`Ex. PGS 1036
`
`
`
`J Elements of3—D Seismology
`tin
`
`pc
`wk
`sc:
`
`=
`=
`
`10 Mflop
`60 Mflop
`
`(1 cpu)
`(1 cpu)
`
`Cray 2
`SGI Power Challenge
`Hitachi SR2201
`
`IBM SP2/402
`Cray T3D MC512-8
`lnte1XP/S
`SGI T3E900
`SGI T3E1200
`lnte1ASC| Red
`
`1.6 Gflop
`110 Gflop
`220 Gflop
`690 Gflop
`510 Gflop
`700 Gflop '
`800 Gflop
`900 Gflop
`1.3 Tflop
`
`(4 cpu)
`(288 cpu)
`(1,024 cpu)
`(402 cpu)
`(512 cpu)
`(3,680 cpu)
`(1,324 cpu)
`(1,084 cpu)
`(9,152 cpu)
`
`=
`=
`=
`=
`=
`=
`
`We should all realize that any list of supercomputer speed is
`extremely volatile. It certainly needs to be updated every 6 months or
`so. The list given here is a mix. Some machines (like the Gray 2) are no
`longer in production. The fastest machine listed (Intel ASCI Red) has
`been online since 1997.
`
`For the latest and greatest information on such things, visit the
`TOP500 web site which aspires to keep an up—to—date list of the world's
`fastest supercomputer installations. The URL is
`http://wvvw.netlib.org/benchmark/top500/top500.list.html
`
`Snead and 3-D Migration
`
`Prestack migration is the classic example of our need for this
`expensive speed. Assume that migration involves 10,000 floating point
`operations per data sample. This is the effort it takes to broadcast each
`input amplitude along an image surface that resembles a lumpy bowl.
`A major part of this work is tied up in calculating the exact geometry of
`the lumpy bowl.
`
`208
`
`EX. PGS 1036
`
`Ex. PGS 1036
`
`
`
`Chapter 77 Computing NeedsJ N(I)
`
`We can estimate how much time is required for 3-D prestack
`depth migration the medium-sized 3-D data set:
`
`Data size:
`
`Assume migration
`total floating point operations
`pc
`=
`wk
`=
`Cray 2
`=
`SGI Power Challange
`=
`SGI T3E1200
`=
`Intel ASCI Red
`=
`
`1,500 samp/trace * 38,400,000 traces
`= 5.76 x10"10 data samples
`= 10,000 operations/sample
`=5.76X10"14
`
`333 days
`90 days
`100 hours
`87 minutes
`9 minutes
`7.4 minutes
`
`(disk?,RAM?,speed?,i/0?,forget it!)
`(disk?,RAM?,speed?,|/0?,good luck!)
`(RAM?,|/O?,dedicated?)
`(RAM?,|/O?,dedicated?)
`(RAM?,l/0?,dedicated?)
`(RAM?,|/0?,dedicated?)
`
`The qualifier "dedicated" is added to highlight the fact that it is
`unusual to have a super computer working on only one problem at a
`time. More likely, jobs are queued up like traffic on a Los Angeles free-
`way, thus slowing work on any given job. There also are issues of disk
`storage, data transfer rates, and grossly insufficient RAM. We should
`therefore consider these estimates extremely optimistic.
`
`
`
`209
`
`EX. PGS 1 036
`
`Ex. PGS 1036