-
How does in-network compute align (if at all) with the E2E principle?
-
Do we have network paradigms for which the benefits of programmable network components outweight the complexity burden?
-
Which problem areas are worth being looked at from a research perspective?
- Slicing
- Machine learning assisted network functions
- Network support for machine learning
- Multicast
- For DLT, diffusion-based multicast would be very useful since its current realization in systems like Ethereum, for instance, is entirely done through (a) unicast replication and (b) an expensive randomized pool mechanisms (done some evaluation work in collaboration with Joerg on this showing the impact on transport networks through such chatty way of doing diffusion multicast).
- Addressing
- Information-Centric Network functions
- Service routing (for IPv6 networks), which is somewhat simpler than ICN
- compute-aware traffic steering (for service routed solutions)
- Security
- Resilience
- what else?
-
Which benefits do we expect from programmable network devices?
- Flexibility (of what?), e.g., flexible placement (of functions), flexible actions, …
- Updatability
- Slicing/multi-tenant support
- what else?
-
What dimensions of complexity do we have?
-
Device Target Hardware Technology
- ASIC (e.g. P4 Tofino)
- FPGA (e.g. NetFPGA)
- Network Processors (e.g. Netronome NFP-6xxx)
- CPUs for Software targets
- what else?
-
Device Target Software Technology
- Compiler tool chain for programmable ASIC (e.g. P4 Tofino) with
- FPGA tool chain (e.g. P4-NetFPGA)
- Network Processors tool chain (e.g. tools for Netronome NFP)
- Software stacks for software targets (e.g. DPDK + P4@ELTE - T4P4S)
- what else?
-
Network Architecture Complexity
- Control plane
- Management planing
- Slicing configuration
- Security functions
- What else?