How do packets cross a changing network?

Follow the same packets through shortest paths, queues, TTL, failure convergence and receiver reordering.

The interactive experiment needs a supported browser.

Preparing the experiment…

0:00Each router chooses a next hop
Follow packet one. Its destination stays R throughout the journey, but each router makes its own next-hop decision. The router receives the complete packet, checks its local table, then sends it onward. The strip below separates processing, transmission and propagation, so the journey time comes from specific operations rather than an animation speed.
0:22Shortest means least routing cost
Look at the numbers beside the links. Shortest path means the lowest sum of routing costs, not the shortest line on the screen. Compare two already-converged networks. Raising only the B-to-R cost sends the packet through C instead. These configured costs do not automatically change whenever a packet joins a queue.
0:44Wait, then serialize the packet
A burst reaches the same output faster than it can transmit. Packet three has to wait while earlier packets are sent. Queueing and serialization are different delays. Each packet here contains twelve hundred bytes, and this output carries four hundred thousand bits per second. Serialization alone takes twenty-four milliseconds, before propagation is added.
1:08TTL bounds forwarding
Give this packet a deliberately small TTL. It leaves the source with two. The first router reduces it to one; the second reduces it to zero and discards it, even though the destination is only one link away. This hop-by-hop bound prevents a packet from circulating indefinitely in a routing loop.
1:30A neighbor notices the failure first
Watch the link from B to the receiver. Packet one is still propagating when that link fails, so it is lost there. B notices first and sends the violet topology update. The other routers must receive and process that message. They cannot all acquire the new network map at the instant of failure.
1:52Routes converge after updates spread
Follow packet two and compare the three local tables. B has changed its route, but C still sends the packet back using old information. That creates a temporary loop. Once C installs the update, it can forward directly to R. The different version numbers expose how convergence happens at different routers at different times.
2:14A full queue drops arriving packets
Send exactly the same ten packets at exactly the same times. With two waiting slots, a full queue discards later arrivals and keeps their IDs in the loss record. Now increase only the buffer. Packet five survives, but waits behind earlier packets. More storage accommodates waiting; it does not create extra link capacity.
2:37Arrival order is not delivery order
Earlier packets are already on the slow route when the faster link returns. Packet four takes that new route and arrives first. The upper row records arrival order. Below it, an example application holds later sequence numbers until gaps are filled. This illustrates ordered delivery; it does not simulate retransmission or a complete TCP connection.

How does a packet reach its destination?

Online messages are divided into small packets, each carrying a destination address. A router acts like a guide at a junction: it checks its own route table and chooses a neighbour.

Each router chooses the next step. Busy exits create queues, broken links can require a detour, and route updates take time to spread.

Follow one packet number. Where does it wait, where does it go, and does it arrive? The same destination does not guarantee the same path or arrival order.

Compare the same packet IDs

Select one packet and move through time to follow its queues and journey.
Break a link. Do nearby and distant route tables update together?
Add waiting space and compare loss with delay. Keeping more packets does not make the exit faster.

Delay, convergence and scope

Further controls and numerical checks

In the stable network, select P3, shorten the release interval and lower the B–R rate. Move time to watch it occupy a waiting slot before being sent.
Set initial TTL to 2. A reduces it to 1; B reduces it to 0 and discards it even though R is nearby.
Select B–R failure and stop at 50 ms. B has changed its route while A and C retain old information. Their updates install at 93–94 ms.
Compare the original buffer with eight waiting slots, keeping the same releases. More packets survive, but the output cannot transmit more bits per second.
Select restoration and reordering, then follow P4. It arrives first, but the example application waits for P1, P2 and P3. No retransmission is simulated: a permanently missing earlier packet leaves a permanent gap.

All six sliders and selection buttons support keyboard operation. Reset restores every parameter, scenario, sample time, packet selection, router selection and buffer comparison. Watch and exploration use the same event model. Millisecond events are slowed for viewing.

Want to check the technical details?

An online message travels in small packets of data. Each packet carries a destination address. A router acts like a signpost at a junction: it checks its own map and chooses the next neighbor, without planning the packet’s entire trip.

Packet one travels from S to R. Routers A and B do not write its entire journey in advance. Each receives the complete packet, looks up a next hop in its own table and selects that output queue. Its destination stays fixed while its path depends on the local state encountered along the way.

Each router in this scene has a link-state database and uses Dijkstra’s algorithm to find the least-cost path. Cost is a configured weight, not the length of a drawn line or an automatic measurement of queueing delay. The initial databases are synchronized. Later, only B–R changes, and other routers must receive and install its update before changing their tables. This implements the central link-state idea, not the complete OSPF protocol. RFC 2328

Loss, temporary loops and reordered arrivals all come from the event model. Every packet keeps its identity and release time, with a record of processing, waiting, transmission, propagation and final outcome.

What contributes to end-to-end delay?

End-to-end delay = Σ(processing + queueing + 8L / R + propagation)

L is bytes and R is bits per second. A 1200 B packet takes 24 ms to serialize at 0.4 Mbit/s. Propagation is measured from the last bit leaving to the last bit arriving. Only then can the next router begin its 1 ms processing step. A packet marker stays at the output during serialization and moves along the link during propagation. It represents packet state, not physical packet dimensions.

Each direction has an independent FIFO output. Waiting slots exclude the packet currently being serialized. With zero waiting slots, an idle output can still immediately transmit a packet. An arriving packet is discarded when the transmitter is busy and all waiting slots are occupied. Packet one in the default network takes 35.2 ms: 2 ms processing, 19.2 ms transmission and 14 ms propagation, with no queueing.

Why do routers learn about failure at different times?

B detects its local B–R failure at 40 ms, updates its own database and originates version 1. An 80 B teaching control message uses a separate 64 kbit/s control budget, then incurs link propagation and 40 ms installation processing at each receiving router. C installs at 93 ms and A at 94 ms. Duplicate versions are neither reinstalled nor reflooded. These deliberately chosen rates and processing times are not OSPF timer values or measured router performance.

Before C updates, B’s new table points to C while C’s old table points back to B. P2 actually follows S → A → B → C → B → C → B → C → R, arriving with TTL=1. A router decrements TTL before forwarding and discards an expired packet; the short processing intervals here use a decrement of one. The destination does not perform another forwarding check. RFC 1812, section 5.3.1

In this model, a failed link discards packets currently transmitting, propagating or waiting on that link. Old scheduled events cannot resurrect them. Exact fault location, detection delay, repair retransmission and protection schemes are excluded. Unchanged links are already known at the start.

Is the receive-order display TCP?

No. P1, P2 and so on are teaching identities and application sequence positions, not IPv4 fragmentation Identification fields. In the restoration scene, the first three packets take C–R with its longer propagation time. Later packets use B–R after routing information changes. The calculated arrival order is P4, P1, P5, P2, P6, P3. The lower row marks only a received contiguous prefix as ready for application delivery; later arrivals are held behind gaps.

IP does not promise reliable delivery or provide end-to-end acknowledgment, retransmission or flow control. RFC 791 This scene also omits TCP windows, acknowledgments, timers and retransmission, fragmentation, ICMP, prefix lookup, ECMP, OSPF adjacency setup, areas and reliable flooding. Green slots illustrate one application ordering policy.

Every planned ID remains in exactly one category: unsent, active, arrived or dropped. Application release is a subset of arrivals and must not be counted again. The simulation never silently replaces losses or removes their IDs from the ledger.

How does a packet reach its destination?

Compare the same packet IDs

Delay, convergence and scope

Further reading & sources

Where will your curiosity go next?