Overview: miscible rpc in coherent multiarch dmaspaces, allows osmotic refactoring

Central concept in various diverse ramblings since systemclick.org started
as pencil sketches in 2013 before I joined Cavium, where the ideas of migration
of workload to programmable offload engines seemed applicable, but go no traction.
Now dropping these disorganized notes to public location preparatory to joining
Google, to avert any future claims that they might have been developed within Google.

Inspired by some beautiful design abstractions in a project that was conceived with
one hand behind it's back, destined never to upstream - the Click Modular Router.

Also Xpra, an X11-py-rpc-py-X11 as a target of a slightly different sister project,
osmotic py rpc

Also inspiration from early access to the no-shared-mem message passing unix port
of the elxi6400, a radical 1980's capability architecture

But I'll now recast it in the language of kernel/eBPF/user interactions, to show how
we can model the expected performance improvement of FPGA-or-ASIC-or-cloud decomposition

eBPF is increasingly being used in places where user/kernel or kernel/hw or user/hw
interactions can be precompiled and prevalidated. This means various interactions
can shortcut more expensive callchains which dance between kernel/user spaces to
consult user policy, when the policy can be formulated in eBPF, validated,
installed in kernel, so less trust/locality zones need to be transited to process
an event.

The same technique can be used where NVMe/comp.stg allows lower layers of
filesystem or search processing to be subcontracted to storage leaves