Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Update Slurm scripts for IB #265

Merged
merged 1 commit into from
Aug 10, 2022
Merged

Update Slurm scripts for IB #265

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
27 changes: 2 additions & 25 deletions gpu_bdb/cluster_configuration/cluster-startup-slurm.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,7 +75,7 @@ if [ "$ROLE" = "SCHEDULER" ]; then
fi

if [ "$CLUSTER_MODE" = "IB" ]; then
DASK_RMM__POOL_SIZE=1GB CUDA_VISIBLE_DEVICES='0' DASK_UCX__CUDA_COPY=True DASK_UCX__TCP=True DASK_UCX__NVLINK=True DASK_UCX__INFINIBAND=True DASK_UCX__RDMACM=False UCX_NET_DEVICES=mlx5_1:1 nohup dask-scheduler --dashboard-address 8787 --protocol ucx --interface ibp18s0 --scheduler-file $SCHEDULER_FILE > $LOGDIR/$HOSTNAME-scheduler.log 2>&1 &
DASK_RMM__POOL_SIZE=1GB CUDA_VISIBLE_DEVICES='0' DASK_DISTRIBUTED__COMM__UCX__CREATE_CUDA_CONTEXT=True nohup dask-scheduler --dashboard-address 8787 --protocol ucx --interface ibp18s0 --scheduler-file $SCHEDULER_FILE > $LOGDIR/$HOSTNAME-scheduler.log 2>&1 &
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Do we also need UCX_MEMTYPE_REG_WHOLE_ALLOC_TYPES and UCX_MAX_RNDV_RAILS with ucx 1.13 or are these enabled by default?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

UCX_MEMTYPE_REG_WHOLE_ALLOC_TYPES=cuda is the default on UCX >= 1.12.0, and UCX_MAX_RNDV_RAILS=1 is the default in UCX-Py since RAPIDS 22.04 or so.

fi

if [ "$CLUSTER_MODE" = "TCP" ]; then
Expand All @@ -90,30 +90,7 @@ if [ "$CLUSTER_MODE" = "NVLINK" ]; then
fi

if [ "$CLUSTER_MODE" = "IB" ]; then
# GPU 0
CUDA_VISIBLE_DEVICES=0 UCX_NET_DEVICES=mlx5_1:1 python -m dask_cuda.cli.dask_cuda_worker --rmm-pool-size $POOL_SIZE --scheduler-file $SCHEDULER_FILE --local-directory $LOCAL_DIRECTORY --interface ibp18s0 --enable-tcp-over-ucx --device-memory-limit $DEVICE_MEMORY_LIMIT --enable-nvlink --enable-infiniband --disable-rdmacm 2>&1 | tee $LOGDIR/$HOSTNAME-worker-0.log &

# GPU 1
CUDA_VISIBLE_DEVICES=1 UCX_NET_DEVICES=mlx5_0:1 python -m dask_cuda.cli.dask_cuda_worker --rmm-pool-size $POOL_SIZE --scheduler-file $SCHEDULER_FILE --local-directory $LOCAL_DIRECTORY --interface ibp12s0 --enable-tcp-over-ucx --device-memory-limit $DEVICE_MEMORY_LIMIT --enable-nvlink --enable-infiniband --disable-rdmacm 2>&1 | tee $LOGDIR/$HOSTNAME-worker-1.log &

# GPU 2
CUDA_VISIBLE_DEVICES=2 UCX_NET_DEVICES=mlx5_3:1 python -m dask_cuda.cli.dask_cuda_worker --rmm-pool-size $POOL_SIZE --scheduler-file $SCHEDULER_FILE --local-directory $LOCAL_DIRECTORY --interface ibp84s0 --enable-tcp-over-ucx --device-memory-limit $DEVICE_MEMORY_LIMIT --enable-nvlink --enable-infiniband --disable-rdmacm 2>&1 | tee $LOGDIR/$HOSTNAME-worker-2.log &

# GPU 3
CUDA_VISIBLE_DEVICES=3 UCX_NET_DEVICES=mlx5_2:1 python -m dask_cuda.cli.dask_cuda_worker --rmm-pool-size $POOL_SIZE --scheduler-file $SCHEDULER_FILE --local-directory $LOCAL_DIRECTORY --interface ibp75s0 --enable-tcp-over-ucx --device-memory-limit $DEVICE_MEMORY_LIMIT --enable-nvlink --enable-infiniband --disable-rdmacm 2>&1 | tee $LOGDIR/$HOSTNAME-worker-3.log &

# GPU 4
CUDA_VISIBLE_DEVICES=4 UCX_NET_DEVICES=mlx5_7:1 python -m dask_cuda.cli.dask_cuda_worker --rmm-pool-size $POOL_SIZE --scheduler-file $SCHEDULER_FILE --local-directory $LOCAL_DIRECTORY --interface ibp148s0 --enable-tcp-over-ucx --device-memory-limit $DEVICE_MEMORY_LIMIT --enable-nvlink --enable-infiniband --disable-rdmacm 2>&1 | tee $LOGDIR/$HOSTNAME-worker-4.log &

# GPU 5
CUDA_VISIBLE_DEVICES=5 UCX_NET_DEVICES=mlx5_6:1 python -m dask_cuda.cli.dask_cuda_worker --rmm-pool-size $POOL_SIZE --scheduler-file $SCHEDULER_FILE --local-directory $LOCAL_DIRECTORY --interface ibp141s0 --enable-tcp-over-ucx --device-memory-limit $DEVICE_MEMORY_LIMIT --enable-nvlink --enable-infiniband --disable-rdmacm 2>&1 | tee $LOGDIR/$HOSTNAME-worker-5.log &

# GPU 6
CUDA_VISIBLE_DEVICES=6 UCX_NET_DEVICES=mlx5_9:1 python -m dask_cuda.cli.dask_cuda_worker --rmm-pool-size $POOL_SIZE --scheduler-file $SCHEDULER_FILE --local-directory $LOCAL_DIRECTORY --interface ibp204s0 --enable-tcp-over-ucx --device-memory-limit $DEVICE_MEMORY_LIMIT --enable-nvlink --enable-infiniband --disable-rdmacm 2>&1 | tee $LOGDIR/$HOSTNAME-worker-6.log &

# GPU 7
CUDA_VISIBLE_DEVICES=7 UCX_NET_DEVICES=mlx5_8:1 python -m dask_cuda.cli.dask_cuda_worker --rmm-pool-size $POOL_SIZE --scheduler-file $SCHEDULER_FILE --local-directory $LOCAL_DIRECTORY --interface ibp186s0 --enable-tcp-over-ucx --device-memory-limit $DEVICE_MEMORY_LIMIT --enable-nvlink --enable-infiniband --disable-rdmacm 2>&1 | tee $LOGDIR/$HOSTNAME-worker-7.log &

python -m dask_cuda.cli.dask_cuda_worker --rmm-pool-size $POOL_SIZE --scheduler-file $SCHEDULER_FILE --local-directory $LOCAL_DIRECTORY --interface ibp18s0 --enable-tcp-over-ucx --device-memory-limit $DEVICE_MEMORY_LIMIT 2>&1 | tee $LOGDIR/$HOSTNAME-worker.log &
fi

if [ "$CLUSTER_MODE" = "TCP" ]; then
Expand Down