[sg0]

I am following your discussions on the GraphBLAS email list. Are you interested in the distributed-memory API of GraphBLAS (for clusters, typically over MPI) or can you make use of it if it were available (you may know, CombBLAS already has a distributed-memory implementation)?

[michelpp]

Yes, I have only reviewed ComBLAS's documentation, but would eventually like to dig into it deeper. A thought I had for distributed graphs would be something like support on top of a distributed store like citus, where shards hold subgraphs. This may be an interesting way to sidestep Postgres' 1GB limit on varlena objects as well.

Edit: I'd forgot to ask, do you have some ideas on the subject? I'm all ears.

[sg0]

Thanks for the clarification. I basically wanted to know if database systems could make use of existing distributed-memory HPC programming models, and it seems that there is a possibility, however engineering efforts might be significant. I really have no experience with database engineering, and I am not sure if MPI and Citus will play out well. I hope you are considering submitting a paper/poster at HPEC'19 for your current work (http://www.ieee-hpec.org/), I think that's a good venue to engage with the GraphBLAS community in person.

[michelpp]

Good idea! Hopefully I can put something together.

[codekilla]

What about something like kdb+/q? I realize it is proprietary software, but it is an array language with super fast vector operations designed to operate on distributed data. I was actually thinking about how well graphBLAS might map to this system. I work in academia, so the non-commercial license is applicable, but I see how it might be an issue for some.