The Port Talbot projectThree agents, three clocks and one merged log
A simulation of agents sending each other messages, and three ways of putting their logs in order.
The agents here are simulated, and the numbers that drive them are our own choices, stated plainly: three agents, something happening somewhere in the system about ten times a second, a little over a third of those events being a message to another agent, and messages taking a fifth of a second on average to arrive. Each agent stamps every event with its own wall clock, and each clock is wrong by a fixed amount, up to the "skew" the learner sets. Twenty runs of about three hundred events produce 1,537 messages to examine.
The first check is the most basic rule of cause and effect: a message cannot be received before it was sent. The learner merges the three logs, sorts them by wall-clock time, and counts how often a receive appears earlier than its send. With perfect clocks the answer is zero. It does not stay zero for long.
Messages whose receipt is stamped before their sending, our Python run| Clocks wrong by up to | Sorted by wall clock | Sorted by Lamport clock |
|---|
| 0 seconds | 0 of 1,537 | 0 |
| 0.05 seconds | 84 (5.5%) | 0 |
| 0.2 seconds | 305 (19.8%) | 0 |
| 0.5 seconds | 484 (31.5%) | 0 |
| 1 second | 618 (40.2%) | 0 |
A Lamport clock is just a counter. Each agent adds one for every event, attaches its counter to every message, and on receiving a message jumps its own counter past the one it was sent. That guarantees the number on a receipt is always bigger than the number on its send, whatever the wall clocks say, which is why the right-hand column stays at zero even when clocks are a full second out.
Lamport clocks have a blind spot, though. Many pairs of events have no connection at all: two agents each did something without hearing from the other. Of 899,102 pairs of events across the twenty runs, 57,104 were concurrent in exactly that sense. Lamport counters still give 52,786 of those pairs different numbers, so a log sorted by them shows an order that never existed. Vector clocks fix this by having each agent carry a small list, one counter per agent. Comparing two lists tells you whether one event came first, the other did, or neither, and in our runs they flagged all 57,104 concurrent pairs correctly. The price is that every message carries a list as long as the number of agents: three numbers here, a hundred in a system of a hundred agents.
Ages 8 to 11
Pass notes around the room with a counter on each one, and rebuild the story of who heard what first.
Ages 11 to 15
Simulate two agents in Python with skewed clocks and catch a reply logged before its question.
Ages 15 and up
Add Lamport and vector clocks, count concurrent pairs, and explain what each clock can and cannot tell you.
Where this comes from
The clock rules follow Lamport (1978); vector clocks are explained in Schwarz and Mattern (1994). The agents, their timings and every count above come from our own simulation; nothing on this page describes a real computer system in Port Talbot.
