It's not trivial, but it's certainly possible. <Un-named IC customer> has been doing level 1 image formation processing for decades using shared memory across a network. They're even doing it in Kubernetes now utilizing libraries from Sandia National Laboratories that use a volume mount of an in-memory filesystem to running containers so you can scale your compute cluster up and down as your memory load varies. This basic setup is powering the entire US geointelligence enterprise. They call it the Tumbo Framework. There's a press release alluding to it here: https://www.sandia.gov/news/publications/labnews/articles/20...

ScaleMP, SGI UV, and HP Superdome all disagree with you. Arguably, plain old RDMA qualifies under this phrasing.

I took the parent comment as meaning on a regular Ethernet+IP network. SGI UV and HP Superdome are semi-exotic hardware architectures. ScaleMP requires Infiniband, which is at least 10 times, if not 100, the price of Ethernet; it also requires fiber optic cable for runs > 5 meters.

I should add to my above comment about the IC doing this kind of thing in Kubernetes on EC2 instances that they utilize the Amazon Elastic Fabric Adapter, which bypasses the ethernet NIC to avoid all of the queueing delay you can't tolerate in applications that need to be basically unaware that some threads are local and some are remote. And obviously they make sure your servers are in the same or physically nearby racks.

I thought modern variations of Ethernet (Gbps+ range) were competing with traditional uses of Infiniband nowadays.

UV and Superdome are custom hardware for huge NUMA boxes,so not a great comparison. ScaleMP is definitely valid though - a real shame it is stuck behind a license fee, would be interesting to experiment with.

Even a hardware RDMA is snail slow for launching a non-adapted piece of software over a network on shared memory.

HPC programs which do use shared memory, were written with network latency in mind from the start.