The Communication SW
Is MPI!
- And the variety will remain
- You need different protocols whether
- running with 5 us HW or 100 us HW
- Scale to 16 or to 10’000
- Sandia says: no connection based protocols to scale
- This is perfectly right, but:
- This is only working if you don’t have a lot of unexpected messages
- Scaling codes don’t have unexpected messages
- And how many codes of our users do have unexpected messages?
- A lot of users don’t have to scale to 10’000 – they are happy with 16
- Software error checking an correction
- Consumes bandwidth and latency
- Can only be an intermediate solution till we have that on network level
- Go and bang on your favorite interconnect vendor!