TCP
Cellular RTT fluctuates greatly, so the timeout and retransmission count should be increased; enable TCP keepalive to detect dead connections, with intervals shorter than the NAT timeout. TLS handshake is expensive, so session reuse and long connections should be maximized.
MQTT
Trade-offs between Clean Session and persistent sessions; QoS1 may be duplicated under weak networks, so consumers need to be idempotent. Coordinate the will message and heartbeat intervals with PSM cycles.
Monitoring
Record the number of connections and the first-packet latency to distinguish between RF issues and cloud-side problems.
Cellular mobility and connection state jitter.
When vehicles and handheld devices switch between cells, TCP may experience brief packet loss and RTT spikes; if misjudged as congestion and the window is overly reduced, application completion time will lengthen. Clients and servers should collaboratively set a more tolerant lower bound for RTO and mark 'first packet after handover' based on RRC state changes reported by cellular modules. MQTT long connections may disconnect during handover and require quick reconnection and evaluation of Clean Session strategy to avoid losing unacknowledged messages. When using dual SIMs or changing carriers, source IP changes may invalidate server sessions, so design should support token renewal and rebinding.
Power-saving modes should coordinate with the protocol stack.
PSM, eDRX, and modem sleep will interrupt TCP window probing and MQTT ping; it is necessary to define an "acceptable offline duration" and a retransmission queue at the application layer. TLS session ticket reuse can reduce handshake overhead after wake-up, but the key update cycle should be balanced with the security policy. For devices that only report occasionally, short connections with HTTP may be more power-efficient than staying always online, and decisions should be based on actual current measurements. Besides RTT, monitoring metrics should include RSRP, SINR, retransmission rate, and module temperature, to establish a baseline for problem localization.
Engineering Practice Supplement
Regarding the implementation of "cellular TCP/MQTT optimization" in IoT and cellular module engineering, it is recommended to list network constraints (target market bands and network sunsetting plans, APN and private network strategies, roaming and tariffs, PSM/eDRX and other power-saving modes) and terminal constraints (antenna and RF layout, power ripple and surge, thermal design and derating, EMC/safety regulations) as verifiable acceptance items at the project initiation stage. During the R&D phase, reproduce boundary failures using weak network simulation, high/low temperatures, and voltage pull deviations, keeping logs and waveforms; for mass production, establish batch and calibration traceability, and handle component replacement through change review. After deployment, include connection success rate, reconnection count, latency, and power consumption on the operations dashboard, with anomalies traceable to versions and batches.
When a Session Still Counts as Alive under High Delay and Handover
Tuning TCP and MQTT on a cellular link means setting timeouts, keepalive, retransmission, and session recovery to the radio round-trip time and the lifetime of address translation, rather than keeping a wired-network default stack. Packet loss that TCP sees may come from a cell handover, not from path congestion. If the window is reduced too far, finishing the job takes much longer. MQTT keeps subscriptions on top of that. With a clean session, a disconnect drops the subscription. With a persistent session, the server may hold unacknowledged messages. Both have to live with the fact that module sleep freezes the link.
This is usable once you know roughly how long the operator idle mapping lasts. Keepalive must be shorter, but not so short that the device cannot enter power saving. With power saving on, the application has to accept a stretch of unreachability and write queue length and the retry cap into the design. The clean-session choice in MQTT has to be made together with that offline interval. A transport-security session ticket can shorten the handshake after wake-up; when the ticket expires, the stack must be able to fall back to a full handshake. If the server binds a session to the source address, a new bearer with a new address must be allowed to rebind, or the client keeps failing to reconnect. Address changes are covered in NAT and private networks.
When keepalive is longer than the mapping timeout, the mapping on the path is already gone while the application still believes it is online, and every downlink command is dropped. At-least-once delivery repeats on a weak network. A consumer without an idempotency key runs a switch command twice. A round-trip spike at handover, treated as congestion, can leave the retransmission timeout elevated for a long time. If unacknowledged messages are not stored before sleep, after a power loss the platform and the device disagree about which one was sent.
Monitoring should separate registration failure, connect failure, handshake failure, and a rejected connection acknowledgement. They are not the same problem. Reconnects need a widening interval and a cap, so a large population does not hit the server at once when the base station returns. For a device that reports only occasionally, compare the charge of a short connection with that of a long one before deciding the session is worth keeping.