-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathresnet50_benchmark.xml
404 lines (379 loc) · 25.2 KB
/
resnet50_benchmark.xml
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
<?xml version="1.0" encoding="UTF-8"?>
<!-- Tensorflow benchmark for
JEDI-GH200 NVIDIA Node:
4× NVIDIA GH200 Grace-Hopper Superchip
CPU: NVIDIA Grace (Arm Neoverse-V2), 72 cores at 3.1 GHz base frequency; 120 GB LPDDR5X memory at 512 GB/s (8532 MHz)
GPU: NVIDIA Hopper H100, 132 multiprocessors, 96 GB HBM3 memory at 4 TB/s
NVIDIA NVLink-C2C CPU-to-GPU link at 900 GB/s
Network: 4× InfiniBand NDR200 (Connect-X7)
TDP: 680 W (for full GH200 superchip)
Available: 44 nodes
JURECA-GH200 NVIDIA:
1× NVIDIA GH200 Grace-Hopper Superchip
CPU: 1 × NVIDIA Grace, 72 cores
GPU: 1 × NVIDIA H100
Memory: 480 GiB LPDDR5X and 96 GiB HBM3
Network: 1 × NVIDIA ConnectX-7 @ 2 × EDR (200 Gbit/s)
Available: 1 chip
JURECA-WAIH100 NVIDIA Node:
4x NVIDIA H100 Hopper-NVLink
CPU: 2x Intel Xeon Platinum 8462Y (Sapphire Rapids) CPUs
A total of 64 computing cores per node à 2.8 GHz (Base Freq.)
Memory: 512 GB DDR5 RAM per node
GPU: 4x NVIDIA H100 GPUs (incl. NVLink Interconnect);94 GB HBM2e per GPU
Available: 16 nodes
JURECA-H100 NVIDIA Node:
1× NVIDIA H100 Hopper-PCIe
CPU: Intel Xeon Platinum Sapphire Rapid 8452Y processor; 2 sockets, 36 cores per socket,\
SMT-2 (total: 2×36×2 = 144 threads) (details for Intel Xeon Platinum 8452Y on Intel ARK)
Memory: 512 GiB DDR5-4800 RAM (of which at least 20 GB is taken by the system software stack,\
including the file system); 256 GB per socket;
GPU: 4 × NVIDIA H100 PCIe GPUs, each with 80 GB memory;
Network: 1 × 1x BlueField-2 ConnectX-6 DPU @ EDR (100 Gbit/s)
Available: 1 node
JURECA-A100 NVIDIA Node:
4x NVIDIA A100 Ampere-SXM
CPU: 2× AMD EPYC 7742, 2× 64 cores, 2.25 GHz
Memory: 512 (16× 32) GB DDR4, 3200 MHz
GPU: 4× NVIDIA A100 GPU, 4× 40 GB HBM2e
Network: 2× InfiniBand HDR (NVIDIA Mellanox Connect-X6)
Available: 192 nodes
JURECA-MI200 AMD Node:
2× AMD MI250
CPU: AMD EPYC 7443 processor (Milan); 2 sockets, 24 cores per socket,
SMT-2 (total: 2×24×2 = 96 threads) in NPS-4 1 configuration (details for AMD EPYC 7443 on WikiChip)
Memory: 512 GiB DDR4-3200 RAM (of which at least 20 GB is taken by the system software stack,
including the file system); 256 GB per socket; 8 memory channels per socket (2 channels per NUMA domain)
GPU: 4 × AMD MI250 GPUs, each with 128 GB memory; the GPUs are built as Multi Chip Modules (MCM)
and because of that they are shown as 8 GPUs with 64 GB memory each.
Network: 1 × Mellanox HDR InfiniBand ConnectX 6 (100 Gbit/s)
Available: 2 nodes
JURECA-M2000 Graphcore Node:
The IPU-POD4 consists of two parts:
AMD EPYC based access server on which user applications are launched with
CPU: AMD EPYC 7413 (Milan); 2 sockets, 24 cores per socket, SMT-2 (total: 2×24×2 = 96 threads) in NPS-4 1 configuration (details for AMD EPYC 7413 on WikiChip)
Memory: 512 GiB DDR4-3200 RAM (of which at least 20 GB is taken by the system software stack, including the file system); 256 GB per socket; 8 memory channels per socket (2 channels per NUMA domain)
Network: 1 × Mellanox EDR InfiniBand ConnectX 5 (100 Gbit/s) to connect to other compute nodes and 1 × Mellanox 100 GigE ConnectX 5 to connect to the IPU-M2000
Graphcore IPU-M2000 which is connected directly to the access server with
IPUs: 4 × GC200 IPUs
Available: 1 node
-->
<jube>
<!-- resnet50_benchmark_run -->
<benchmark name="tensorflow_benchmark" outpath="resnet50_benchmark_run">
<parameterset name="systemInfo">
<parameter name="system_version" mode="shell">echo 2024.01</parameter>
<parameter name="system_name" type="str" tag="JEDI">JEDI</parameter>
<parameter name="system_name" type="str" tag="GH200">GH200</parameter>
<parameter name="system_name" type="str" tag="WAIH100">WAIH100</parameter>
<parameter name="system_name" type="str" tag="H100">H100</parameter>
<parameter name="system_name" type="str" tag="A100">A100</parameter>
<parameter name="system_name" type="str" tag="MI250">MI250</parameter>
<parameter name="system_name" type="str" tag="GC200">GC200</parameter>
</parameterset>
<parameterset name="systemParameter" init_with="platform.xml">
<parameter name="nodes" type="int" separator=";" tag="!container" mode="python">
{
"JEDI": "1;2",
"GH200": "1",
"WAIH100": "1",
"H100": "1" ,
"A100": "1;2",
"MI250": "1;2",
"GC200": "1"
}["${system_name}"]</parameter>
<parameter name="nodes" type="int" separator=";" tag="container+(H100|GH200|MI250|GC200)" mode="python">
{
"GH200": "1",
"H100": "1" ,
"MI250": "1",
"GC200": "1"
}["${system_name}"]</parameter>
<parameter name="taskspernode" type="int" separator=";" tag="!container" mode="python">
{
"JEDI": "1;2;3;4",
"GH200": "1",
"WAIH100": "1;2;3;4",
"H100": "1;2;3;4" ,
"A100": "1;2;3;4",
"MI250": "1;2;3;4;5;6;7;8",
"GC200": "1"
}["${system_name}"]</parameter>
<parameter name="taskspernode" type="int" separator=";" tag="container+(H100|GH200|MI250|GC200)" mode="python">
{
"GH200": "1",
"H100": "1" ,
"MI250": "1",
"GC200": "1"
}["${system_name}"]</parameter>
<parameter name="tasks" type="int" separator=";" mode="python">${nodes}*${taskspernode}</parameter>
<parameter name="threadspertask" type="int" tag="!container" mode="python">
{
"JEDI": "72",
"GH200": "72",
"WAIH100": "16",
"H100": "4" ,
"A100": "4",
"MI250": "4",
"GC200": "12"
}["${system_name}"]</parameter>
<parameter name="threadspertask" type="int" tag="container" mode="python">4</parameter>
<parameter name="queue" mode="python">
{
"JEDI": "all",
"GH200": "dc-gh",
"WAIH100": "dc-wai",
"H100": "dc-h100" ,
"A100": "dc-gpu",
"MI250": "dc-mi200",
"GC200": "dc-ipu"
}["${system_name}"]</parameter>
<parameter name="account" tag="!WAIH100">zam</parameter>
<parameter name="account" tag="WAIH100">westai0005</parameter>
<parameter name="timelimit" tag="!GC200+!container">00:20:00</parameter>
<parameter name="timelimit" tag="GC200+!container">01:20:00</parameter>
<parameter name="timelimit" tag="(H100|GH200|MI250|GC200)+container">01:20:00</parameter>
<!-- When using Horovod ntasks=hvd_size=num_workers; there can be only one gpu per worker; hence num_gpus=1 -->
<parameter name="gpus" type="int" separator=";">1</parameter>
<parameter name="singularity_file" mode="python">
{
<!-- https://docs.nvidia.com/deeplearning/frameworks/tensorflow-release-notes/rel-23-01.html -->
"JEDI": "$root_dir/containers/ngc2301_tf115_cuda1201_nccl2165_py38_arm.sif",
"GH200": "$root_dir/containers/ngc2301_tf115_cuda1201_nccl2165_py38_arm.sif",
<!-- https://docs.nvidia.com/deeplearning/frameworks/tensorflow-release-notes/rel-23-01.html -->
"WAIH100": "$root_dir/containers/ngc2301_tf115_cuda1201_nccl2165_py38.sif",
"H100": "$root_dir/containers/ngc2301_tf115_cuda1201_nccl2165_py38.sif" ,
"A100": "$root_dir/containers/ngc2301_tf115_cuda1201_nccl2165_py38.sif",
<!-- https://hub.docker.com/layers/rocm/tensorflow/rocm5.0-tf2.7-dev/images/sha256-664fbd3e38234f5b4419aa54b2b81664495ed0a9715465678f2bc14ea4b7ae16?context=explore -->
"MI250": "$root_dir/containers/amd_tf27_rocm50_py39-dev.sif",
<!-- https://hub.docker.com/layers/john856/caraml/tensorflow_poplar310_tf263_mpi4py/images/sha256-57cb664cb1e3493657c576b07a0d274363e1097e62820d2f7e03db5e68fe1f0e?context=repo -->
"GC200": "$root_dir/containers/ipu_tf263_poplar310_py38.sif"
}["${system_name}"]</parameter>
<!-- Benchmark file -->
<parameter name="executable" tag="!GC200+!container">python ${folder}/tf_cnn_benchmarks.py</parameter>
<parameter name="executable" tag="GC200+!container">python3 ${folder}/train.py</parameter>
<parameter name="executable" tag="(H100|GH200|MI250|GC200)+container">bash ${jube_benchmark_home}/get_tensorflow_container.sh</parameter>
<!-- Benchmark file arguments -->
<parameter name="args_exec" separator="!" tag="GC200+!container"> ${data_arg} --config=${tf_model} --validation=false --mlperf-logging --num-replicas=${replicas} --num-epochs=1 --global-batch-size=${global_batch_size} --micro-batch-size=${ipu_mbs} --dbn-replica-group-size=${dbn_replica}</parameter>
<parameter name="args_exec" separator="!" tag="!GC200+!container+!synthetic">--num_gpus $gpus --model $tf_model --batch_size ${batch_size_per_device} --data_dir=${data_dir} --variable_update horovod --use_fp16=True</parameter>
<parameter name="args_exec" separator="!" tag="!GC200+!container+synthetic">--num_gpus $gpus --model $tf_model --batch_size ${batch_size_per_device} --variable_update horovod --use_fp16=True</parameter>
<parameter name="measurement" separator="!">${load_modules}; time -p</parameter>
</parameterset>
<parameterset name="modelParameter">
<parameter name="tf_model" tag="!GC200">resnet50_v2</parameter>
<parameter name="tf_model" tag="GC200">resnet50_mlperf_pod4_bs20</parameter>
<!-- Data Parallel for IPU -->
<!-- 1;2;3;4 -->
<parameter name="replicas" type="int" separator=";" tag="GC200">1;2;3;4</parameter>
<parameter name="total_devices" type="int" separator=";" mode="python" tag="!GC200">${nodes}*${taskspernode}</parameter>
<parameter name="total_devices" type="int" separator=";" mode="python" tag="GC200">${replicas}</parameter>
<!--Replace with path to ImageNet data -->
<parameter name="data_dir" tag="!synthetic">/p/project1/cjsc/benchmark/imagenet-processed/</parameter>
<parameter name="data_arg" tag="GC200+!synthetic">--dataset-path=${data_dir}</parameter>
<parameter name="data_arg" tag="GC200+synthetic">--synthetic-data=ipu</parameter>
<parameter name="dataset" tag="synthetic">ImageNet(synthetic)</parameter>
<parameter name="dataset" tag="!synthetic">ImageNet</parameter>
<!-- 16;32;64;128;256;512;1024;2048;4096 -->
<parameter name="global_batch_size" type="int" separator=";">16;32;64;128;256;512;1024;2048;4096</parameter>
<parameter name="dbn_replica" type="int" separator=";" tag="GC200">${replicas}</parameter>
<parameter name="batch_size_per_device" type="int" separator=";" mode="python" tag="!GC200">int(${global_batch_size}/${total_devices})</parameter>
<parameter name="batch_size_per_device" type="int" separator=";" mode="python" tag="GC200">int(${global_batch_size}/${replicas})</parameter>
<!-- For IPU make sure the global batch size = replicas*mbs*gradient_accumulation_count -->
<!-- ref: https://github.com/chelseajohn/examples/blob/d6f1846182e0a6edad4b79898aa60b973fe8732e/vision/cnns/tensorflow2/batch_config.py#L30 -->
<parameter name="ipu_mbs" mode="python" type="int" tag ="GC200" separator="!">$batch_size_per_device if ($global_batch_size == 16) or ($global_batch_size == 32 and $replicas == 4) else 16</parameter>
<!-- IPU hacky workaround: JUBE does not go to next step due to error134 -->
<parameter name="ipu_energy_file" tag="GC200+!container">$jube_benchmark_home/$jube_wp_relpath/GC200_power.0.energy.csv</parameter>
</parameterset>
<parameterset name="environment">
<parameter name="root_dir">$jube_benchmark_home/..</parameter>
<parameter name="ipu_log_flag" separator="!">export POPLAR_LOG_LEVEL=INFO;export POPART_LOG_LEVEL=INFO</parameter>
<parameter name="ipu_cache_path" seperator="!">export TF_POPLAR_FLAGS='--executable_cache_path=$jube_benchmark_home/ipu_cache'</parameter>
<parameter name="python_path">export PYTHONPATH=${packages}:$PYTHONPATH</parameter>
<parameter name="bench_path">export BENCH_DIR=${jube_benchmark_home}</parameter>
<parameter name="data_path" tag="!synthetic">export DATASETS_DIR=${data_dir}</parameter>
<parameter name="mpi_modules">GCC OpenMPI</parameter>
<parameter name="accelerator">export ACCELERATOR=$system_name</parameter>
<parameter name="cuda_devices" separator="!">export CUDA_VISIBLE_DEVICES=0,1,2,3</parameter>
<parameter name="rocm_devices" separator="!">export ROCM_VISIBLE_DEVICES=0,1,2,3,4,5,6,7</parameter>
<parameter name="xla" separator="!">export TF_XLA_FLAGS='--tf_xla_auto_jit=2'</parameter>
<parameter name="packages" separator="!" mode="python">
{
<!-- nvidia_arm_tensorflow_requirements.txt -->
"JEDI": "$jube_benchmark_home/nvidia_arm_tensorflow_packages/lib/python3.8/site-packages",
"GH200": "$jube_benchmark_home/nvidia_arm_tensorflow_packages/lib/python3.8/site-packages",
<!-- nvidia_x86_tensorflow_requirements.txt -->
"WAIH100": "$jube_benchmark_home/nvidia_x86_tensorflow_packages/lib/python3.8/site-packages",
"H100": "$jube_benchmark_home/nvidia_x86_tensorflow_packages/lib/python3.8/site-packages" ,
"A100": "$jube_benchmark_home/nvidia_x86_tensorflow_packages/lib/python3.8/site-packages",
<!-- amd_requirements.txt -->
"MI250": "$jube_benchmark_home/amd_tensorflow_packages/lib/python3.9/site-packages",
<!-- ipu_requirements.txt -->
"GC200": "$jube_benchmark_home/ipu_tensorflow_packages/lib/python3.8/site-packages"
}["${system_name}"]</parameter>
<parameter name="load_modules" separator="!" tag="!container" mode="python">
{
"JEDI": "${cuda_devices}; ${python_path};${accelerator}; ${bench_path}",
"GH200": "${cuda_devices}; ${python_path};${accelerator}; ${bench_path}",
"WAIH100": "${cuda_devices}; ${python_path};${accelerator}; ${bench_path}",
"H100": "${cuda_devices}; ${python_path};${accelerator}; ${bench_path}" ,
"A100": "${cuda_devices}; ${python_path};${accelerator}; ${bench_path}",
"MI250": "${xla}; ${rocm_devices}; ${python_path};${accelerator}; ${bench_path}",
"GC200": "${ipu_log_flag}; ${ipu_cache_path}; ${accelerator}; ${bench_path}"
}["${system_name}"]</parameter>
<parameter name="load_modules" separator="!" tag="container" mode="python">
{
"GH200": "${accelerator}; ${bench_path}",
"H100": "${accelerator}; ${bench_path}" ,
"MI250": "${accelerator}; ${bench_path}",
"GC200": "${accelerator}; ${bench_path}"
}["${system_name}"]</parameter>
</parameterset>
<parameterset name="executeset" init_with="platform.xml">
<parameter name="folder" tag="!GC200">$jube_benchmark_home/$jube_wp_relpath/benchmarks/scripts/tf_cnn_benchmarks</parameter>
<parameter name="folder" tag="GC200">$jube_benchmark_home/$jube_wp_relpath/graphcore_benchmarks/vision/cnns/tensorflow2</parameter>
<parameter name="bind_dir" tag="!container+!synthetic" separator="!">$jube_benchmark_home,${folder},${data_dir}</parameter>
<parameter name="bind_dir" tag="!container+synthetic" separator="!">$jube_benchmark_home,${folder}</parameter>
<parameter name="args_starter" separator="!" tag="!container" mode="python">
{
"JEDI": " --cpu_bind=socket,v --mpi=pmi2 apptainer exec --bind=${bind_dir} --nv ${singularity_file} ${jube_benchmark_home}/${jube_wp_relpath}/nvidia_arm_tensorflow_wrap.sh ",
"GH200": "--cpu_bind=socket,v --mpi=pspmix env PMIX_SECURITY_MODE=native apptainer exec --bind=${bind_dir} --nv ${singularity_file} ${jube_benchmark_home}/${jube_wp_relpath}/nvidia_arm_tensorflow_wrap.sh ",
"WAIH100": "--cpu_bind=none,v --mpi=pspmix env PMIX_SECURITY_MODE=native apptainer exec --bind=${bind_dir} --nv ${singularity_file} ${jube_benchmark_home}/${jube_wp_relpath}/nvidia_x86_tensorflow_wrap.sh ",
"H100": "--cpu_bind=none,v --mpi=pspmix env PMIX_SECURITY_MODE=native apptainer exec --bind=${bind_dir} --nv ${singularity_file} ${jube_benchmark_home}/${jube_wp_relpath}/nvidia_x86_tensorflow_wrap.sh " ,
"A100": "--cpu_bind=none,v --mpi=pspmix env PMIX_SECURITY_MODE=native apptainer exec --bind=${bind_dir} --nv ${singularity_file} ${jube_benchmark_home}/${jube_wp_relpath}/nvidia_x86_tensorflow_wrap.sh ",
"MI250": "--cpu_bind=none,v --mpi=pspmix env PMIX_SECURITY_MODE=native apptainer exec --bind=${bind_dir} ${singularity_file} ${jube_benchmark_home}/${jube_wp_relpath}/amd_tensorflow_wrap.sh ",
"GC200": "--cpu_bind=none,v --mpi=pspmix env PMIX_SECURITY_MODE=native apptainer exec --bind=${bind_dir} ${singularity_file} ${jube_benchmark_home}/${jube_wp_relpath}/ipu_tensorflow_wrap.sh "
}["${system_name}"]</parameter>
</parameterset>
<fileset name="files" tag="!GC200">
<prepare>git clone -b master --recursive https://github.com/chelseajohn/tf_cnn_benchmarks.git benchmarks</prepare>
</fileset>
<fileset name="files" tag="GC200">
<prepare>git clone -b master --recursive https://github.com/chelseajohn/examples.git graphcore_benchmarks</prepare>
</fileset>
<step name="pull_container" tag="container+(H100|GH200|MI250|GC200)" iterations="1">
<use>systemInfo</use>
<use>systemParameter</use>
<use>environment</use>
<use>executeset</use>
<use from="platform.xml">jobfiles</use>
<use from="platform.xml">executesub</use>
<do>$submit $submit_script</do>
<do done_file="$done_file"></do>
</step>
<step name="execute" tag="!container" iterations="1">
<use>files</use>
<use>systemInfo</use>
<use>systemParameter</use>
<use>modelParameter</use>
<use>environment</use>
<use>executeset</use>
<use from="platform.xml">jobfiles</use>
<use from="platform.xml">executesub</use>
<do mode="shell" tag="GC200+synthetic">printf "%s\n" "export PYTHONPATH=${packages}:\$PYTHONPATH" "\$*" > ${jube_benchmark_home}/${jube_wp_relpath}/ipu_tensorflow_wrap.sh</do>
<do mode="shell" tag="(H100|A100|WAIH100)+synthetic">printf "%s\n" "export PYTHONPATH=${packages}:\$PYTHONPATH" "\$*" > ${jube_benchmark_home}/${jube_wp_relpath}/nvidia_x86_tensorflow_wrap.sh</do>
<do mode="shell" tag="(GH200|JEDI)+synthetic">printf "%s\n" "export PYTHONPATH=${packages}:\$PYTHONPATH" "\$*" > ${jube_benchmark_home}/${jube_wp_relpath}/nvidia_arm_tensorflow_wrap.sh</do>
<do mode="shell" tag="MI250+synthetic">printf "%s\n" "export PYTHONPATH=${packages}:\$PYTHONPATH" "\$*" > ${jube_benchmark_home}/${jube_wp_relpath}/amd_tensorflow_wrap.sh</do>
<do mode="shell" tag="GC200+!synthetic">printf "%s\n" "export PYTHONPATH=${packages}:\$PYTHONPATH" "${data_path}" "\$*" > ${jube_benchmark_home}/${jube_wp_relpath}/ipu_tensorflow_wrap.sh</do>
<do mode="shell" tag="(H100|A100|WAIH100)+!synthetic">printf "%s\n" "export PYTHONPATH=${packages}:\$PYTHONPATH" "${data_path}" "\$*" > ${jube_benchmark_home}/${jube_wp_relpath}/nvidia_x86_tensorflow_wrap.sh</do>
<do mode="shell" tag="(GH200|JEDI)+!synthetic">printf "%s\n" "export PYTHONPATH=${packages}:\$PYTHONPATH" "${data_path}" "\$*" > ${jube_benchmark_home}/${jube_wp_relpath}/nvidia_arm_tensorflow_wrap.sh</do>
<do mode="shell" tag="MI250+!synthetic">printf "%s\n" "export PYTHONPATH=${packages}:\$PYTHONPATH" "${data_path}" "\$*" > ${jube_benchmark_home}/${jube_wp_relpath}/amd_tensorflow_wrap.sh</do>
<do mode="shell" tag="GC200"> chmod a+rwx ${jube_benchmark_home}/${jube_wp_relpath}/ipu_tensorflow_wrap.sh </do>
<do mode="shell" tag="(H100|A100|WAIH100)"> chmod a+rwx ${jube_benchmark_home}/${jube_wp_relpath}/nvidia_x86_tensorflow_wrap.sh </do>
<do mode="shell" tag="(GH200|JEDI)"> chmod a+rwx ${jube_benchmark_home}/${jube_wp_relpath}/nvidia_arm_tensorflow_wrap.sh </do>
<do mode="shell" tag="MI250"> chmod a+rwx ${jube_benchmark_home}/${jube_wp_relpath}/amd_tensorflow_wrap.sh </do>
<do>$submit $submit_script</do>
<do done_file="$done_file"></do>
</step>
<step name="combine_energy" tag="!container+!GC200" depend="execute" iterations="1">
<use>files</use>
<use>systemInfo</use>
<use>systemParameter</use>
<use>modelParameter</use>
<do>module load GCC SciPy-bundle</do>
<do>python3 ${jube_benchmark_home}/aux/combine_energy.py --output ${jube_benchmark_home}/${jube_wp_relpath}/combined_energy.csv $$(printf "${jube_benchmark_home}/${jube_wp_relpath}/execute/${system_name}_power.%d.energy.csv " $$(seq 0 $$((${tasks}-1))))</do>
</step>>
<patternset name="pattern">
<pattern name="imagespersec" type="float" tag="!GC200">total images/sec: ${jube_pat_fp}</pattern>
<pattern name="imagespersec" type="float" tag="GC200">throughput: ${jube_pat_fp}</pattern>
<pattern name="gas" type="int" tag="GC200">INFO:root:gradient accumulation ${jube_pat_int}</pattern>
<pattern name="mbs_per_exec" type="int" tag="GC200">INFO:root:micro_batches_per_execution = ${jube_pat_int}</pattern>
<pattern name="compilation_time" type="float" unit="s" tag="GC200">logging_callback:Compilation Time ${jube_pat_fp}</pattern>
<!-- Assuming 90 epochs with 1,281,167 training samples of ImageNet, with the simple formula
[ time_to_report_in_seconds ] = [epochs] * [images] / [throughput in images/second] -->
<!-- <pattern name="throughput_in_time" type="float" mode="python" unit="s">(90 * 1281167 /$imagespersec)</pattern> -->
<!-- <pattern name="energy_average" tag="GC200">Average-E
nergy-per-GPU: ${jube_pat_fp}</pattern> -->
<!-- <pattern name="energy_list_reimann" type="str" >Energy-per-GPU-list integrated\(Wh\): (\[.*?\])</pattern> -->
<!-- <pattern name="energy_list_counter" type="str" tag="MI250" >Energy-per-GPU-list from counter\(Wh\): (\[.*?\])</pattern> -->
</patternset>
<patternset name="efile_patterns">
<pattern name="energy_file_path" type="str" tag="!GC200">Writing combined energy DataFrame to (.*)</pattern>
</patternset>
<patternset name="jobnumber">
<pattern name="jobid">Submitted batch job $jube_pat_int</pattern>
</patternset>
<patternset name="runtimepattern">
<pattern name="runtime" type="float">real $jube_pat_fp</pattern>
<pattern name="error_code" type="int">JUBE_ERR_CODE=$jube_pat_int</pattern>
</patternset>
<analyser name="analyse" reduce="false">
<analyse step="combine_energy" tag="!GC200">
<file use="efile_patterns">stdout</file>
</analyse>
<analyse step="execute">
<file use="pattern">job.out</file>
<file use="pattern,runtimepattern">job.err</file>
<file use="pattern,jobnumber">stdout</file>
</analyse>
</analyser>
<result>
<use>analyse</use>
<table name="result" style="pretty" sort="total_devices,nodes">
<column title="JobID">jobid</column>
<column title="System">system_name</column>
<column title="Version">system_version</column>
<column title="Queue">queue</column>
<column title="Runtime(s)">runtime</column>
<column title="Model">tf_model</column>
<column title="Dataset">dataset</column>
<column title="Nodes">nodes</column>
<column title="Devices">total_devices</column>
<column title="Tasks/Node">taskspernode</column>
<column title="Threads/Task">threadspertask</column>
<column title="GlobalBatchSize">global_batch_size</column>
<column title="BatchSize/Device">batch_size_per_device</column>
<column title="Images/second" format=".2f">imagespersec_avg</column>
<!-- <column format=".2f">throughput_in_time</column> -->
<!-- <column title="Energy/Device(Wh)">energy_list_reimann</column> -->
<!-- <column title="Energy_counter/Device(Wh)" tag="MI250">energy_list_counter</column> -->
<!-- <column title="Energy/Device(Wh)" format=".2f">energy_average</column> -->
<column title="EnergyFile" tag="!GC200">energy_file_path</column>
<column title="EnergyFile" tag="GC200">ipu_energy_file</column>
<column>args_starter</column>
<column>args_exec</column>
</table>
<table name="result_csv" style="csv" sort="total_devices,nodes">
<column title="JobID">jobid</column>
<column title="System">system_name</column>
<column title="Version">system_version</column>
<column title="Queue">queue</column>
<column title="Runtime(s)">runtime</column>
<column title="Model">tf_model</column>
<column title="Dataset">dataset</column>
<column title="Nodes">nodes</column>
<column title="Devices">total_devices</column>
<column title="Tasks/Node">taskspernode</column>
<column title="Threads/Task">threadspertask</column>
<column title="GlobalBatchSize">global_batch_size</column>
<column title="BatchSize/Device">batch_size_per_device</column>
<column title="Images/second" format=".2f">imagespersec_avg</column>
<!-- <column format=".2f">throughput_in_time</column> -->
<!-- <column title="Energy/Device(Wh)">energy_list_reimann</column> -->
<!-- <column title="Energy_counter/Device(Wh)" tag="MI250">energy_list_counter</column> -->
<column title="EnergyFile" tag="!GC200">energy_file_path</column>
<column title="EnergyFile" tag="GC200">ipu_energy_file</column>
<!-- <column title="Energy/Device(Wh)" format=".2f">energy_average</column> -->
</table>
</result>
</benchmark>
</jube>