Network Bound Dgraph Cluster Topologies, Data Sets or Queries

[rcgoodfellow (Ryan Goodfellow)]

Hello Dgraph Community,

I am a researcher working in data center network infrastructure. My current aim is to identify data center workloads that are network i/o bound on today’s networks, and deploy them on new network substrate technologies that are being developed in the research community to assess the potential impact.

I have done some simple tests using the throughput test in the Dgraph benchmark repository on a topology composed on 1 zero and 4 nodes. The data was sharded 4 ways before hand using the bulk loader. I’ve also taken the liberty of making the queries a bit more complex than the original benchmark by expanding connections out a bit further. However, initial results show that with this query/data/topology combination - Dgraph is not terribly i/o bound (10 gbps links to the nodes seem to be more than sufficient).

My question is if any one in the community has run across situations in which Dgraph is network i/o bound?

Thanks!

~ ry

[mrjn (Manish R Jain)]

The community might have some interesting situations. But, from a design perspective, it would be hard to network bound Dgraph. The design is such that we decrease the network communication required to execute queries. Furthermore, we also transmit data in integer format, so the payload sizes are small.