CUDA Quantum Docs Bot: A preview of the documentation can be found here.

Please correct me if I'm wrong but my understanding is that this variable applies to the communication network between all MPI processes regardless of whether they are intra-node or inter-node. For example, the given list would also apply to 4 nodes and 8 GPUs/node, with NVLinks between the GPUs. The current formulation makes the impression it only applies to the network between nodes and the that the network inside a node is not relevant. Here is a suggestion that reflects my understanding:

Specify the network structure (faster to slower). For example, assuming a 32 MPI processes simulation, whereby the network topology is divided into 4 groups of 8 processes, which have faster communication network between them. In this case, the CUDAQ_GLOBAL_INDEX_BITS environment variable can be set to 3,2. The first 3 (log2(8)) represents 8 processes with fast communication within the group and the second 2 represents the 4 groups (8 processes each) in those total 32 processes. The sum of all elements in this list is 5, corresponding to the total number of MPI processes (2^5 = 32). Default is an empty list (no customization based on network structure of the cluster).

Please correct me if I'm wrong but my understanding is that this variable applies to the communication network between all MPI processes regardless of whether they are intra-node or inter-node. For example, the given list would also apply to 4 nodes and 8 GPUs/node, with NVLinks between the GPUs. The current formulation makes the impression it only applies to the network between nodes and the that the network inside a node is not relevant. Here is a suggestion that reflects my understanding:

Specify the network structure (faster to slower). For example, assuming a 32 MPI processes simulation, whereby the network topology is divided into 4 groups of 8 processes, which have faster communication network between them. In this case, the CUDAQ_GLOBAL_INDEX_BITS environment variable can be set to 3,2. The first 3 (log2(8)) represents 8 processes with fast communication within the group and the second 2 represents the 4 groups (8 processes each) in those total 32 processes. The sum of all elements in this list is 5, corresponding to the total number of MPI processes (2^5 = 32). Default is an empty list (no customization based on network structure of the cluster).

That makes sense. I've incorporated your suggestions in 26e3801

CUDA Quantum Docs Bot: A preview of the documentation can be found here.