This. I'm really bummed that so many of these things-- OpenMosix, OpenSSI, Kerrighed, etc, have died.
There's a lot of utility in stitching together many systems into one-- a middle ground between loosely coupled application-controlled distributed concurrency and tightly coupled SMP/NUMA.
Like if you want a big multiuser build farm, for instance, where you're not drowning in IPC and synchronization, but not everything maps well to distcc.
For what it's worth, ScaleMP will still sell you something more-or-less like that, at least for HPC systems, but I've never used it, and it seemed to have significant restrictions. There's also still (proprietary) bproc: https://www.penguincomputing.com/documentation/scyld-cluster...
Could be interesting. Some kind of opensource zOS parallel sysplex or zVM single system image. Shame that the last update was done in 2010. Any other known alternatives?
I remember back in 2006, when multicore processors were rare and multisocket servers were expensive, how cool and powerful our OpenSSI cluster seemed to me as a junior sysadmin. Technology really went in a different direction, though, and we're buuilding new abstractions that bridge machines (e.g. kubernetes) instead of trying to extend unix across them.
I never used something like OpenSSI, but I'd be surprised if running over (presumably) Ethernet of the time worked well for an SSI compared with the "expensive" things like SGI Origin with fast links and hardware coherence. Solutions like that still exist, and there's the SCI standard. Current things like NVlink and CAPI seem to be rather in the same direction.
I don't think Kubernetes is a fundamentally different abstraction to the (HPC) distributed resource managers that ran on shared- and distributed-memory systems of the time. (In that space MPI somewhat standardizes remote memory access and process spawning.)
9 comments
[ 2.5 ms ] story [ 31.2 ms ] threadThere's a lot of utility in stitching together many systems into one-- a middle ground between loosely coupled application-controlled distributed concurrency and tightly coupled SMP/NUMA.
Like if you want a big multiuser build farm, for instance, where you're not drowning in IPC and synchronization, but not everything maps well to distcc.
I don't think Kubernetes is a fundamentally different abstraction to the (HPC) distributed resource managers that ran on shared- and distributed-memory systems of the time. (In that space MPI somewhat standardizes remote memory access and process spawning.)