2SYN: Dual Use in TCP and Quantum Memory

• In networking, 2SYN is a TCP multihoming algorithm that races duplicate SYN packets to select the lowest-delay path without requiring end-host cooperation.
• In neuromorphic quantum materials, 2SYN refers to a second-order synaptic memory in tDBLG, where the state is encoded via nonlinear electrical responses rather than conventional conductance.
• Both implementations showcase innovative design: one optimizes connection setup with minimal overhead, while the other achieves multilevel synaptic states without traditional charge-trapping mechanisms.

The label 2SYN appears in current arXiv literature in two distinct technical senses. In computer networking, it denotes a congestion-aware multihoming algorithm for TCP that selects a WAN path for each new connection by racing duplicate SYN packets and committing the flow to the path that returns the first SYN-ACK (Toledo et al., 27 Aug 2025). In quantum and neuromorphic-device research, it denotes a second-order synaptic memory realized in twisted double bilayer graphene, where the programmable state is read out through a second-order nonlinear electrical response rather than only through linear conductance (Ahmed et al., 1 Jun 2026). These usages are unrelated and belong to different research communities.

1. Name, scope, and domain-specific meanings

The networking usage is the only one that appears as an explicit paper title, namely "2SYN: Congestion-Aware Multihoming" (Toledo et al., 27 Aug 2025). The condensed-matter usage appears as an acronymic shorthand for second-order synaptic memory, introduced for a dual-gated, hBN-encapsulated AB–AB stacked twisted double bilayer graphene device (Ahmed et al., 1 Jun 2026).

Usage of “2SYN”	Domain	Defining meaning
2SYN	Networking	TCP multihoming via duplicate SYN transmission and first-SYN-ACK path selection
2SYN	Neuromorphic quantum materials	Memory encoded in second-order nonlinear electrical response in tDBLG

In the networking sense, 2SYN addresses enterprise multihoming in which a branch router has multiple WAN links such as LTE, DSL, cable, or fiber and must choose a path for outbound TCP flows to arbitrary public destinations, including previously unseen servers (Toledo et al., 27 Aug 2025). In the neuromorphic sense, 2SYN addresses the realization of synaptic functionality in a single moiré quantum material without charge-trapping layers, ferroelectric gate stacks, ionic motion, or phase-change media (Ahmed et al., 1 Jun 2026).

2. 2SYN as congestion-aware multihoming

In networking, 2SYN is defined as a congestion-aware multihoming algorithm for TCP that works for any destination, including previously unseen ones (Toledo et al., 27 Aug 2025). Its central design constraint is that the router cannot assume remote cooperation, cannot require MPTCP support, and cannot rely on an SD-WAN endpoint, tunnel, or cooperating gateway at the far side. The destination may be any public Internet server.

The mechanism exploits a universally available TCP signal: the 3-way handshake. For a new TCP flow, identified by its 5-tuple, the multihoming router duplicates the outbound SYN across all available WAN links. Because each WAN link has its own public-facing source address, the router performs NAT so that each duplicate SYN uses the IP address of the corresponding outgoing WAN interface. The destination therefore sees multiple ordinary connection attempts, one per WAN path (Toledo et al., 27 Aug 2025).

The router then waits for the first SYN-ACK. The path on which that SYN-ACK arrives is taken to be the preferred path for the connection, corresponding to the path with the shortest initial RTT. Once the winner is known, the router installs per-flow routing state, records the choice in an Active Flow (AF) table, and pins the connection to that WAN link for its lifetime. Duplicate connection attempts on losing path(s) are cancelled by sending a RST to the destination (Toledo et al., 27 Aug 2025).

The workflow is therefore connection-establishment-centric rather than throughput-estimation-centric. It uses only locally visible handshake packets—SYN, SYN-ACK, FIN, and RST—and it does not depend on historical measurements, active probes, or destination-specific training. This is what allows it to operate for arbitrary TCP destinations, including those with no prior traffic history (Toledo et al., 27 Aug 2025).

3. Implementation model, evaluation, and operational behavior

The Linux proof-of-concept combines iptables for packet classification and SYN duplication, Linux routing tables for per-flow path selection, NAT for source rewriting, and Python Scapy with BPF filters for monitoring only control packets of interest (Toledo et al., 27 Aug 2025). On SYN arrival, the router marks the flow as awaiting its first SYN-ACK; on first SYN-ACK arrival, it adds the flow to the AF table and installs the corresponding routing-table entry; on FIN or RST, it removes the state from both tables.

The reported overhead is modest in the prototype. Adding or removing a route takes about 4 ms on average; AF-table operations are effectively $O(1)$; and when processing a burst of 100 parallel SYN arrivals, peak CPU overhead was about 3% (Toledo et al., 27 Aug 2025). The traffic overhead per new TCP flow is one additional SYN, one additional SYN-ACK, and one RST for the non-selected path. The destination-side half-open-connection overhead is estimated as roughly

$RTT / FCT_D$

and, under the paper’s assumption that median $FCT_D$ is about 100 RTTs based on a median flow duration around 6 s, this is estimated around 1% (Toledo et al., 27 Aug 2025).

The evaluation uses both a controlled lab and real LTE-versus-wired experiments. The lab testbed employs five Ubuntu 20.04 Linux servers, iPerf3 3.9, and NetEm. Baselines are: always choose path 1, always choose path 2, choose randomly with equal probabilities, and 2SYN (Toledo et al., 27 Aug 2025). In one experiment with both paths at 300 Mbps but RTTs of 120 ms and 80 ms, 2SYN always selected the lower-delay path for 1 MB downloads repeated 20 times. In another, both paths had RTT 120 ms and bandwidth 300 Mbps, but one path carried background TCP traffic of five flows, each limited to 100 Mbps; 2SYN chose the less congested path and achieved flow-completion times close to the better fixed-path baseline (Toledo et al., 27 Aug 2025).

The real wide-area study places the source at the Technion in Israel and uses public SpeedTest servers in England. The two links are path 1, a standard 4G LTE link via cellphone, and path 2, a wired link throttled to 100 Mbps download / 10 Mbps upload (Toledo et al., 27 Aug 2025). Workloads include 30 MB files, sent 20 times, and a web-search traffic mix consisting of 62% small files $<100\text{ KB}$, 18% medium files $100\text{ KB} - 1\text{ MB}$, and 20% large files $>1\text{ MB}$. Across “no issues,” bandwidth-drop, and congestion-increase scenarios, 2SYN switched paths when the DSL-like link degraded to 5 Mbps or when ten TCP flows were injected as cross-traffic, and it outperformed fixed-path baselines in the representative cases reported (Toledo et al., 27 Aug 2025).

4. Formal properties, comparative position, and limitations

2SYN is positioned against three classes of alternatives: static failover, static load balancing, and history-based learning (Toledo et al., 27 Aug 2025). Static failover prefers one link until failure; static load balancing distributes flows randomly, round-robin, uniformly, or by weights; and lightweight machine-learning baselines are modeled as multi-armed bandits, specifically $\epsilon$-greedy, UCB, and Thompson sampling. Under fixed path conditions, the bandit methods eventually converge to near-optimal behavior, but when conditions change—such as a path-2 bandwidth drop from 300 Mbps to 100 Mbps after 40% of files—2SYN adapts immediately at connection setup and clearly outperforms the MAB approaches (Toledo et al., 27 Aug 2025).

The paper also states a limited optimality theorem. If one path has strictly lower RTT for every packet/ACK pair than all other paths, and packets and ACKs experience no reordering on any path, then always using that path is optimal and 2SYN is also optimal, achieving the same lowest flow completion time (Toledo et al., 27 Aug 2025). A corollary extends this to the case where one path has strictly lower round-trip propagation time and all paths have the same bottleneck capacities with negligible queueing delay.

The main limitations are explicit. First, 2SYN optimizes only at connection establishment. Because WAN links typically imply different source IP addresses, ordinary TCP semantics forbid path changes mid-connection; current flows cannot be migrated if the chosen path degrades after the handshake (Toledo et al., 27 Aug 2025). Second, the first SYN-ACK is only a proxy for eventual flow completion time. The paper notes that this may hurt when a higher-RTT path later provides higher throughput, especially for large flows. Third, the design is TCP-specific. The paper explicitly discusses difficulties for QUIC and notes that for UDP or other non-handshake protocols, duplicating the first packet would duplicate application data (Toledo et al., 27 Aug 2025). Fourth, scalability to large $k$ is conceptually straightforward but incurs one more duplicate SYN and one more half-open attempt per additional path; the experimental focus is therefore on the common case $k=2$.

A common misconception is to equate 2SYN with MPTCP or SD-WAN overlays. The paper is explicit that it requires no changes at the end hosts or applications, no destination cooperation, and no remote tunnel endpoint (Toledo et al., 27 Aug 2025). Its novelty lies precisely in remaining compatible with ordinary TCP and arbitrary Internet destinations.

5. 2SYN as second-order synaptic memory

A separate usage of 2SYN denotes second-order synaptic memory in a single moiré quantum material, twisted double bilayer graphene (tDBLG) (Ahmed et al., 1 Jun 2026). Here the device is a dual-gated, hBN-encapsulated AB–AB stacked tDBLG in which two properties coexist: electronic plasticity / hysteretic memory under vertical displacement field and a second-order nonlinear electrical response whose sign and magnitude are strongly tunable by carrier concentration and displacement field.

The defining distinction from conventional electronic synapses is that the memory state is encoded in the second-order electrical output, not only in first-order conductance. The nonlinear transport expansion is written as

$$j_i = \sigma_{ij} E_j + \chi_{ijk} E_j E_k + \cdots,$$

with the second-order current density

$RTT / FCT_D$0

Experimentally, the device is read out through longitudinal and transverse second-harmonic voltages,

$RTT / FCT_D$1

with the measured relation

$RTT / FCT_D$2

verified for drive currents from 30 nA to 300 nA (Ahmed et al., 1 Jun 2026).

The memory arises from $RTT / FCT_D$3-dependent hysteresis in sheet resistance and nonlinear response, especially near integer moiré fillings. The authors argue against interfacial charge traps and conventional ferroelectricity and instead propose a mechanism based on intrinsic strain-disorder-related metastability, with flexoelectricity associated with strain and strain gradients in low-angle tDBLG providing the more compelling explanation (Ahmed et al., 1 Jun 2026). Hysteresis appears in low-angle devices, is absent in the high-angle control sample at $RTT / FCT_D$4, and persists up to 100 K.

Programming is performed using triangular displacement-field pulses. A fixed negative reset pulse at

$RTT / FCT_D$5

is followed by a positive set pulse varied typically from

$RTT / FCT_D$6

at a sweep rate of

$RTT / FCT_D$7

At a read point around

$RTT / FCT_D$8

different pulse histories yield different metastable values of $RTT / FCT_D$9 and $FCT_D$0, which the authors interpret as potentiation and depression in the second-order signal (Ahmed et al., 1 Jun 2026).

The reported figures of merit are unusual for a nonlinear synaptic element: more than 16 multilevel synaptic states, endurance over $FCT_D$1 cycles, retention with decay as low as $FCT_D$2 per day, and an estimated synaptic-event energy

$FCT_D$3

at

$FCT_D$4

The broader significance is that 2SYN in this sense is a programmable nonlinear susceptibility element, not merely a programmable resistor (Ahmed et al., 1 Jun 2026).

6. Terminological ambiguity and adjacent labels

Because the two explicit uses of 2SYN are unrelated, the term is intrinsically ambiguous outside context. In networking, it refers to a two-SYN handshake-racing procedure for multihomed TCP routing (Toledo et al., 27 Aug 2025). In moiré-device physics, it abbreviates second-order synaptic memory (Ahmed et al., 1 Jun 2026). The overlap is lexical rather than conceptual.

There are also nearby labels that can be confused with 2SYN in search or OCR pipelines. The astronomy paper "The Synchronous Network of distant Telescopes" uses the acronym SNT, not 2SYN, even though the underlying concept is a synchronized multi-site observational system with GPS-based UTC alignment to 1 microsecond and better (Zhilyaev et al., 2011). The paper explicitly notes that an apparent “2SYN” query may arise from a terminology mismatch or acronym confusion rather than from the actual system name. This makes contextual disambiguation essential: a networking query centered on TCP, SYN packets, NAT, or WAN links almost certainly concerns the multihoming algorithm, whereas a materials query involving tDBLG, $FCT_D$5, hysteresis, or moiré superlattices concerns the second-order synaptic memory.

In present arXiv usage, then, 2SYN is not a single unified technical doctrine. It is a homonymous label spanning at least a router-level path-selection mechanism and a moiré-material synaptic device, each defined by a distinct higher-order selection or readout principle and each anchored in its own specialized literature (Toledo et al., 27 Aug 2025, Ahmed et al., 1 Jun 2026).