3D DISCRETE DISLOCATION DYNAMICS APPLIED TO ... - NUMODIS

Recommendations

Info

70 Parallelization of the Discrete Dislocation Dynamics method 3.3.2 Long-distance stresses computations The serial version computes long-distance stresses as follows. The boxes that are situated at long- distance relative to one give box are recognized by a topological relation. The stresses due to the segments in the long-distance labeled boxes are computed at the center point of the given box. Thus one processor scans all the boxes and computes the long-distance stresses in the serial program. In a parallel program, the work is divided into several processors, since each processor is responsible for a fraction of the boxes. Boxes in each processor are distinguished by the array ibs(6), and each processor computes the long-distance stresses of boxes only in its subsystem. The serial and the parallel version are compared in the following. Serial version DO iz=1, M DO iy=1, M DO ix=1, M the box ’ib’ compute the box index ’ib’ compute the long-distance stresses of ENDDO ENDDO ENDDO ⇒ Parallel version DO iz=ibs(5), ibs(6) DO iy=ibs(3), ibs(4) DO ix=ibs(1), ibs(2) compute the box index ’ib’ compute the long-distance stresses of the box ’ib’ ENDDO ENDDO ENDDO The parallel version uses most of the serial codings and only the range of the loops are slightly modified. And it should be noted that each subsystem shares all the segments information at the time of computing the long-distance stresses. 3.3.3 Short-distance stresses computation The following pseudo-code explains how the short-distance stresses are computed both in the serial and in the parallel DDD code.
3.3 Parallelization of the serial DDD program 71 Serial version DO is=1, Nsegm segment ’is’ Identify the box ’ib’ containing the Compute the short-distance stresses due to the segments within the short-distance boxes ENDDO ⇒ Parallel version DO is=1, iscnt(p) Identify the box ’ib’ containing the segment ’is’ Compute the short-distance stresses due to the segments within the short-distance boxes ENDDO As for the computation of the long-distance stresses, only small modifications are made to the serial coding: in the case of the serial program, all the segments (Nsegm) are processed by a single processor. In the parallel program, on the contrary, segments are distributed among several processors, and a processor p computes stresses of iscnt(p) segments only. The construction of iscnt(p) will be discussed in Sec. 3.3.4. Since the stress on a segment can be computed without regard to the stress on the other segments, all processors can work independently. The elapsed time for stress computation decreases by a factor of 1/P (the number of processors), if the number of segments of each processor is the same. Otherwise, the overall elapsed time for the stress computation is determined by the busiest processor, because the other processors have to wait until the latest processor finishes the computation to move the segments. For higher efficiency, the segments have to be distributed uniformly over the different processors. This can be realized by shifting the subsystem boundaries, which changes the ibs array and consequently iscnt. This load balancing issue will be addressed in Sec. 3.4.4. 3.3.4 Data structures for distributing and the gathering segments The processors do not work entirely independently in a parallel program. At some point of a program, it needs to collect all the information to one processor or to distribute the data to all the processors. An obvious example of gathering information is when data are written in external files. One processor normally takes charge of writing files, and the data to be written are sent to that processor from the other processors. In a parallel DDD program, segments’ information including coordinates, neighbors, linked-list and the effective stress etc. need to be communicated. The segments are identified by a vector of integer numbers. To send segments’ data to the other processor, the list of segments to be sent should be shared between the sender and the receiver processors. The arrays iswork(:,:) and iscnt(;) are
Page 1 and 2:
INSTITUT NATIONAL POLYTECHNIQUE DE
Page 3:
Acknowledgements First of all, I ex
Page 6 and 7:
Résumé La dynamique des dislocati
Page 8 and 9:
viii CONTENTS 2.3.5 Plastic strain
Page 11 and 12:
Chapter 1 Introduction 1.1 Computat
Page 13 and 14:
1.2 Dislocation dynamics 3 (a) Weak
Page 15 and 16:
1.3 Scope of Thesis 5 Recently the
Page 17:
1.3 Scope of Thesis 7 coupling meth
Page 20 and 21:
10 Description of the simulation me
Page 22 and 23:
Page 24 and 25:
Page 26 and 27:
Page 28 and 29:
Page 30 and 31: 20 Description of the simulation me
Page 61 and 62: Chapter 3 Parallelization of the Di
Page 63 and 64: 3.1 An introduction to Supercomputi
Page 71 and 72: 3.2 Towards a parallel DDD code 61
Page 77 and 78: 3.3 Parallelization of the serial D
Page 79: 3.3 Parallelization of the serial D
Page 91 and 92: 3.4 Performance improvment 81 Speed
Page 97 and 98: 3.4 Performance improvment 87 Y Z I
Page 99 and 100: 3.5 Application to Stage I-II trans
Page 105 and 106: Chapter 4 Dislocation-precipitate i
Page 107 and 108: 4.1 Image stresses due to a 3D part
Page 115 and 116: 4.2 A simple case of dislocation-pa
Page 127 and 128: 4.3 Fatigue simulations of material
Page 129 and 130: 4.3 Fatigue simulations of material
Page 131 and 132:
4.3 Fatigue simulations of material
Page 133 and 134:
Page 135 and 136:
Page 137 and 138:
Page 139 and 140:
Page 141 and 142:
Page 143 and 144:
Page 145 and 146:
Page 147 and 148:
Page 149 and 150:
Page 151 and 152:
Page 153 and 154:
Page 155 and 156:
Page 157 and 158:
Page 159 and 160:
Chapter 5 Conclusions and perspecti
Page 161 and 162:
namics codes into a parallel versio
Page 163 and 164:
Bibliography [Abraham 97] Abraham F
Page 165 and 166:
BIBLIOGRAPHY 155 [Devincre & Robert
Page 167 and 168:
BIBLIOGRAPHY 157 [Khraishi et al. 0
Page 169 and 170:
BIBLIOGRAPHY 159 [Risbet et al. 03]
show all

3D DISCRETE DISLOCATION DYNAMICS APPLIED TO ... - NUMODIS

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?