Can Multipath TCP Congestion Control do better without adhering to its First Design Principle?

MultiPath TCP (MPTCP) is guided by three classical design principles: ( P1 ) an MPTCP flow should perform at least as well as a TCP flow on its best subflow path; ( P2 ) it should be fair to TCP flows; and ( P3 ) it should enable resource pooling across subflow paths. These principles were intended to ensure that MPTCP can achieve significant throughput improvement over single-path TCP while remaining fair to TCP traffic. However, extensive empirical and experimental studies consistently report that the throughput improvement achieved by MPTCP over TCP is often modest, and the root cause of this limitation remains poorly understood. In this paper, we show that this limitation is not merely an artifact of specific congestion control algorithms (CCAs) or parameter choices. Rather, the joint enforcement of P1 -- P3 under end-to-end transport induces a structural throughput limitation that constrains the achievable MPTCP throughput to remain close to that of the best-path TCP. To further understand and overcome this limitation, we adopt the network utility maximization (NUM) framework to characterize multipath rate allocations and examine how their optimality properties interact with the classical MPTCP design principles. Our analysis shows that relaxing P1 while preserving P2 and P3 admits NUM-optimal allocations that achieve higher throughput than those satisfying all three principles. This demonstrates that relaxing P1 fundamentally enlarges the space of achievable rate allocations, providing a principled avenue for overcoming the throughput limitation imposed by P1 -- P3. In particular, for a very special network setting, the NUM analysis yields a closed-form achievable gain over best path TCP-Reno, which approaches 4-2√2 (approximately 17.2%) as the number of subflows increases, and hence serves as a performance lower bound for the achievable gain under general network settings. Notably, this bound represents a conservative baseline. Our analysis further indicates that even with a slightly more aggressive TCP CCA than TCP-Reno, rate allocations can substantially magnify the gain, pushing throughput improvements well beyond 70% in a large number of path regimes. Guided by this analysis, we design N-MPTCP( → ω), an MPTCP CCA framework that realizes the corresponding NUM-optimal rate allocations. NS-3 simulations corroborate the theoretical predictions, demonstrating that N-MPTCP( → ω) consistently achieves throughput gains aligned with the NUM analysis.