NetLoom: Accelerating the Construction of Large-Scale Network Emulation on a Single Host
TL;DR
This work revisit VN construction from a kernel-execution perspective, and presents NetLoom, a pipelined network emulation construction framework that decouples VN construction into active execution and blocking phases, co-optimizing them through two complementary techniques: overlapped scheduling to mitigate pipeline starvation, and kernel-level pruning to accelerate execution along the synchronous control path.
Abstract
Container-based network emulation has emerged as a scalable approach for high-fidelity network experiments. However, existing emulators spend a disproportionate amount of time constructing the virtual network (VN) rather than executing experiments, severely degrading efficiency. This bottleneck arises primarily from the substantial overhead of virtual device instantiation and the time-consuming, synchronous kernel execution paths for network configuration. To address this challenge, we revisit VN construction from a kernel-execution perspective. Our analysis shows that the conventional staged workflow strictly adheres to time-consuming kernel synchronization constraints, overlooking the in-kernel opportunities to accelerate the construction. Motivated by these insights, we present NetLoom, a pipelined network emulation construction framework. It decouples VN construction into active execution and blocking phases, co-optimizing them through two complementary techniques: overlapped scheduling to mitigate pipeline starvation, and kernel-level pruning to accelerate execution along the synchronous control path. Experimental results on two representative large-scale topologies show that NetLoom achieves a 4.06–7.03× speedup in setup time across the evaluated scales, compared to the conventional staged workflow on a single host.