This commit is contained in:
Fred Park 2016-09-09 14:30:28 -07:00
Родитель 7254e217ec
Коммит f16e185520
6 изменённых файлов: 16 добавлений и 9 удалений

Просмотреть файл

@ -1,7 +1,7 @@
{
"pool_specification": {
"id": "<your pool id>",
"vm_size": "<vm size>",
"vm_size": "STANDARD_F1",
"vm_count": 3,
"inter_node_communication_enabled": true,
"publisher": "Canonical",

Просмотреть файл

@ -69,7 +69,8 @@ For this example, we will run the Multigpu Simple2d Example in the
`alfpark/cntk:gpu-openmpi` Docker image. The application `command` to run
would be:
`"mpirun --allow-run-as-root --host $AZ_BATCH_HOST_LIST --mca btl_tcp_if_exclude docker0 /bin/bash -c \"export LD_LIBRARY_PATH=/usr/local/openblas/lib:/usr/local/nvidia/lib64 && cp -r /cntk/Examples/Other/Simple2d/* . && /cntk/build/gpu/release/bin/cntk configFile=Config/Multigpu.cntk RootDir=. parallelTrain=true\""`
* **NOTE:** tasks that span multiple compute nodes will need their output
* **NOTE:** tasks that span multiple compute nodes
(i.e., MultiNode+SingleGPU or MultiNode+MultiGPU) will need their output
stored on a shared file system, otherwise CNTK will fail during test
as all of the output is written by rank 0 to the specified output
directory only on that compute node. To override the output directory for

Просмотреть файл

@ -11,11 +11,12 @@
"shared_data_volumes": [
"<azure file shared volume name>"
],
"command": "mpirun --allow-run-as-root --host $AZ_BATCH_HOST_LIST,$AZ_BATCH_HOST_LIST,$AZ_BATCH_HOST_LIST,$AZ_BATCH_HOST_LIST --mca btl_tcp_if_exclude docker0 /bin/bash -c \"cp -r /cntk/Examples/Other/Simple2d/* . && /cntk/build/gpu/release/bin/cntk configFile=Config/Multigpu.cntk RootDir=. OutputDir=$AZ_BATCH_NODE_SHARED_DIR/azurefileshare/Output parallelTrain=true\"",
"command": "mpirun --allow-run-as-root --host $AZ_BATCH_HOST_LIST,$AZ_BATCH_HOST_LIST,$AZ_BATCH_HOST_LIST,$AZ_BATCH_HOST_LIST --mca btl_tcp_if_exclude docker0 /bin/bash -c \"export LD_LIBRARY_PATH=/usr/local/openblas/lib:/usr/local/nvidia/lib64 && cp -r /cntk/Examples/Other/Simple2d/* . && /cntk/build/gpu/release/bin/cntk configFile=Config/Multigpu.cntk RootDir=. OutputDir=$AZ_BATCH_NODE_SHARED_DIR/azurefileshare/Output parallelTrain=true\"",
"multi_instance": {
"num_instances": "pool_specification_vm_count",
"coordination_command": null
}
},
"gpu": true
}
]
}

Просмотреть файл

@ -8,11 +8,12 @@
"image": "alfpark/cntk:gpu-openmpi",
"name": "cntk",
"remove_container_after_exit": true,
"command": "mpirun --allow-run-as-root --host $AZ_BATCH_HOST_LIST,$AZ_BATCH_HOST_LIST --mca btl_tcp_if_exclude docker0 /bin/bash -c \"cp -r /cntk/Examples/Other/Simple2d/* . && /cntk/build/gpu/release/bin/cntk configFile=Config/Multigpu.cntk RootDir=. OutputDir=$AZ_BATCH_NODE_SHARED_DIR/azurefileshare/Output parallelTrain=true\"",
"command": "mpirun --allow-run-as-root --host $AZ_BATCH_HOST_LIST,$AZ_BATCH_HOST_LIST --mca btl_tcp_if_exclude docker0 /bin/bash -c \"export LD_LIBRARY_PATH=/usr/local/openblas/lib:/usr/local/nvidia/lib64 && cp -r /cntk/Examples/Other/Simple2d/* . && /cntk/build/gpu/release/bin/cntk configFile=Config/Multigpu.cntk RootDir=. parallelTrain=true\"",
"multi_instance": {
"num_instances": "pool_specification_vm_count",
"coordination_command": null
}
},
"gpu": true
}
]
}

Просмотреть файл

@ -7,7 +7,8 @@
"image": "alfpark/cntk:gpu-openmpi-mnist-cifar",
"remove_container_after_exit": true,
"command": "/bin/bash -c \"cp -r /cntk/Examples/Image/MNIST/* . && /cntk/build/gpu/release/bin/cntk configFile=Config/02_Convolution_ndl_deprecated.cntk RootDir=.\""
}
},
"gpu": true
}
]
}

Просмотреть файл

@ -1,8 +1,11 @@
# Batch Shipyard Recipes
This directory contains recipes and sample batch-style Docker workloads for
use with Batch Shipyard on Azure Batch. Please note that all recipes have
sample configurations that you can use to quickly get going. However, as they
are samples only, you will need to configure them to your liking for actual
sample configurations that you can use to quickly get going, however, some
of the config files cannot be used as-is as they need input such as
the pool `id` in `pool_specification`. Please review each config file you
are planning on using and modify accordingly. As the config samples are bare
examples only, you will need to configure them to your liking for actual
workloads.
**NOTE: Not all recipes are populated.**