HPC@Green IT (eBook)

Ralf Gruber won the Cray Gigaflop Performance Award in 1989 with world’s fastest parallel program running at 1.7 GFlop/s sustained. He was responsible for the Swiss-Tx cluster project, a co-operation between EPFL, Compaq, and Supercomputing Systems. Since 6 years he teaches the doctoral school course on "High Performance Computing Methods".Vincent Keller received his Master degree in Computer Science from the University of Geneva (Switzerland) in 2004, and his PhD degree in 2008 from the Swiss Federal Institute of Technology (EPFL) in the HPCN and HPC Grids fields. Since 2009, Dr. Vincent Keller holds a full-time researcher position at University of Bonn in Germany. His research interests are in HPC applications analysis, Grid and cluster computing and energy efficiency of large computing ecosystems.

Foreword 5
Preface 7
Acknowledgements 9
Contents 11
1 Introduction 16
Basic goals of the book 16
What do I get for one Watt today ? 16
Main memory bottleneck 18
Optimize resource usage 18
Application design 19
Organization of the book 19
Historical aspects 19
Parameterization 20
Models 20
Core optimization 21
Node optimization 21
Cluster optimization 21
Grid-brokering to save energy 22
2 Historical highlights 23
Evolution of computing 23
The first computer companies 28
ERA, EMCC and Univac 28
Control Data Corporation, CDC 28
Cray Research 29
Thinking Machines Corporation 30
International Business Machines (IBM) 31
The ASCI effort 32
The Japanese efforts 33
The computer generations 34
The evolution in computing performance 34
Performance/price evolution 36
Evolution of basic software 36
Evolution of algorithmic complexity 37
The TOP500 list 39
Outlook with the TOP500 curves 41
The GREEN500 List 42
Proposal for a REAL500 list 44
3 Parameterization 45
Definitions 45
Parameterization of applications 49
Application parameter set 49
Parameterization of BLAS library routines 50
SMXV: Parameterization of sparse matrix*vector operation 52
Parameterization of a computational nodes Pi ri 53
Parameterization of the interconnection networks 55
Types of networks 55
Parameterization of clusters and networks 56
Parameters related to running applications 58
Conclusion 61
4 Models 62
The performance prediction model 62
The execution time evaluation model (ETEM) 66
A network performance model 66
The extended -- model 68
Validation of the models 69
Methodology 69
Example: The full matrix*matrix multiplication DGEMM 70
Example: Sparse matrix*vector multiplication SMXV 72
5 Core optimization 75
Some useful notions 75
Data hierarchy 75
Data representation 76
Floating point operations 79
Pipelining 80
Single core optimization 82
Single core architectures 82
Memory conflicts 82
Indirect addressing 86
Unrolling 87
Dependency 88
Inlining 90
If statement in a loop 90
Code porting aspects 91
How to develop application software 95
Application to plasma physics codes 96
Tokamaks and Stellerators 96
Optimization of VMEC 100
Optimization of TERPSICHORE 103
Conclusions for single core optimization 106
6 Node optimization 107
Shared memory computer architectures 107
SMP/NUMA architectures 107
The Cell 111
GPGPU for HPC 112
Node comparison and OpenMP 117
Race condition with OpenMP 121
Application optimization with OpenMP: the 3DHelmholtz solver 122
Fast Helmholtz solver for parallelepipedic geometries 123
NEC SX-5 reference benchmark 125
Single processor benchmarks 126
Parallelization with OpenMP 127
Parallelization with MPI 127
Conclusion 131
Application optimization with OpenMP: TERPSICHORE 131
7 Cluster optimization 133
Introduction on parallelization 133
Internode communication networks 133
Network architectures 133
Comparison between network architectures 141
Distributed memory parallel computer architectures 143
Integrated parallel computer architectures 143
Commodity cluster architectures 146
Energy consumption issues 148
The issue of resilience 149
Type of parallel applications 150
Embarrassingly parallel applications 150
Applications with point-to-point communications 150
Applications with multicast communication needs 151
Shared memory applications (OpenMP) 151
Components based applications 151
Domain decomposition techniques 151
Test example: The Gyrotron 152
The geometry and the mesh 154
Connectivity conditions 154
Parallel matrix solver 155
The electrostatic precipitator 157
Scheduling of parallel applications 158
Static scheduling 158
Dynamic scheduling 158
SpecuLOOS 159
Introduction 159
Test case description 159
Complexity on one node 161
Wrong complexity on the Blue Gene/L 162
Fine results on the Blue Gene/L 163
Conclusions 163
TERPSICHORE 165
Parallelization of the LEMan code with MPI and OpenMP 166
Introduction 166
Parallelization 166
CPU time results 168
Conclusions 171
8 Grid-level Brokering to save energy 173
About Grid resource brokering 173
An Introduction to ïanos 174
Job Submission Scenario 176
The cost model 177
Mathematical formulation 177
CPU costs Ke 179
License fees K 181
Costs due to waiting time Kw 181
Energy costs Keco 181
Data transfer costs Kd 183
Example: The Pleiades clusters CPU cost per hour 183
Different currencies in a Grid environment 185
The implementation 185
Architecture & Design
The Grid Adapter 186
The Meta Scheduling Service (MSS) 187
The Resource Broker 188
The System Information 189
The Data Warehouse 189
The Monitoring Service 189
The Monitoring Module VAMOS 190
Integration with UNICORE Grid System 191
Scheduling algorithm 191
User Interfaces to the ïanos framework 193
DVS-able processors 194
Power consumption of a CPU 195
An algorithm to save energy 196
First results with SMXV 197
A first implementation 198
Conclusions 200
9 Recommendations 201
Application oriented recommendations 201
Code development 201
Code validation 201
Porting codes 202
Optimizing parallelized applications 202
Race condition 202
Hardware and basic software aspects 203
Basic software 203
Choice of system software 204
Energy reduction 204
Processor frequency adaptation 204
Improved cooling 205
Choice of optimal resources 205
Best choice of new computer 205
Last but not least 206
Miscellaneous 206
Course material 206
A new REAL500 List 206
Glossary 208
References 215
About the authors 222
Index 224