jzhao.xyz

Search IconIcon to open search

 .
  \_.
  /

Clocks

Last updated May 5, 2022

Measuring time in the context of computer systems

# Physical Time

Two types of clock

  1. Physical clock: number of seconds elapsed
  2. Logical clock: count events, e.g. messages sent

Time is hard! So many different ways of measuring time

Periodically adjust local clock with a server with a more accurate time source using Network Time Protocol (NTP) or Precision Time Protocol (PTP)

How do we estimate time over a network?

# Logical Time

Problem is that even with synced clocks we can have $t_2 < t_1$ with a message A at $t_1$ with a response B at $t_2$. Here, the timestamp order is inconsistent with expected order! This can happen when the clock skew is less than the one way network latency.

So we use logical clocks to work based off of the number of events that have occurred rather than actual time passed.

# Lamport Clocks

Logic

Properties

This means that two identical Lamport timestamps might not correspond to the same unique event. However if we include the node $N(e)$ for the node where event $e$ occurred, then $(L(e), N(e))$ uniquely identifies event $e$.

We attempt to define a total causal order

$$(a \prec b) \iff (L(a) < L(b)) \lor (L(a) = L(b) \land N(a) < N(b))$$

However even now, given timestamps $L(a) < L(b)$, we can’t tell whether $a \rightarrow b$ or $a \parallel b$

To separate causality from concurrent events, we need vector clocks!

# Vector Clocks

Instead of having a single counter t for all nodes, we keep a vector timestamp $a$ of an event for each node so we have $V(a) = \langle t_1, t_2, \ldots, t_n \rangle$ where $t_i$ is the number of events observed by node $N_i$

Each node has a current vector timestamp $T$, on an event on node $N_i$, increment vector element $T[i]$

Logic

Thus, a vector timestamp of an event $e$ actually represents all of its causal dependencies: ${ e } \cup {a | a \rightarrow e }$

E.g. $\langle 2, 2, 0 \rangle$ represents first two events from $N_1$, first two events from $N_2$, and no events from $N_3$

Ordering

Properties


Interactive Graph