Question on giant graphs

Questions
#1

Currently, only shared-memory graph store server is supported, so store_type
can only be “shared_memory”.

I look into the code of the graph_store.py and found that the server only supports shared memory.

For giant graphs, is NUMA necessary? We prefer K8S or Yarn to schedule our tasks. Any distributed solution for giant graphs (over 100 million nodes)?

References

  1. Distributed or Parallel Training Docs/Guide
  2. Dose DGL support distributed training & inference?
#2

Hi,

100 Million nodes can fit in a machine with large memory, such as x1 instance at aws with 4TB memory. Based on our experiment, NUMA could help performance.