SepaNet: Towards Practical Large-Scale Full-System Simulation for Systems Research#

Hejing Li (MPI-SWS); Marvin Meiers (Saarland University); Jialin Li (NUS); Antoine Kaufmann (MPI-SWS)

• Why: Their premise matches what DC practitioners know: bottlenecks today are in host software and hardware, not the wire protocol, yet most network evaluation still simulates protocols. SepaNet makes end-to-end (full-system) simulation at cluster scale tractable — a methodology shift for anyone reasoning about DC system design.
• What: A simulation framework that lets users navigate fidelity/time/resource tradeoffs explicitly by mixing simulation models of different fidelity with controlled parallelization, guided by a profiler that identifies bottlenecks and informs partitioning. Demonstrated: a 1200-node datacenter network running end-to-end applications, simulated in 175 minutes for 20 s of system time on 24 cores (from abstract only).
• How: User-specified system abstractions instantiate multiple simulations of the same system at different fidelity levels; the framework composes them and parallelizes with awareness of where fidelity actually matters.
• Where applicable: Research and pre-deployment evaluation of distributed/DC systems where physical testbeds at scale are out of reach. A research framework rather than product, but the mixed-fidelity idea is directly borrowable for internal evaluation tooling.