Problem

Currently, dstack uses the default slow inter-node networking in GCP, i.e. 10GB/s on a3 instances. The networking speed can be increased significantly for GPU instances. This requires setting up multiple NICs per VM and configuring RDMA:

• GPUDirect-TCPX for a3 high and edge
• GPUDirect-TCPXO for a3 mega

Solution

No response

Workaround

No response

Would you like to help us implement this feature by sending a PR?

Yes