Lecture 12: Synchronization

Required for correctness and fairness. Challenging because:

Clock Skew	Clock Drift
Relative difference in clock values of 2 processes
Non-zero clock skew = clocks not synchronized	Relative difference in clock frequencies (rates) of 2 processes
Non-zero clock drift = cases skew to increase overtime

MDR (maximum drift rate) between 2 clocks with similar MDR = 2 * MDR
Given maximum acceptable skew between 2 clocks, need to sync at least once every M / (2 * MDR) time units

External Sync	Internal Sync
This is what we cover in this class. External clock with highly precise atomic clock is accepted as source of truth.	Not discussed in this class. Ex: Berkeley’s algorithm.
Internal sync does not imply external sync.

Time Server S (highly precise atomic clock) ← all Processes P sync with S
Algorithm accounts for message latencies
- min1 = P to S (ex: client → network interface → server)
- min2 = S to P (ex: server → network interface → server)
- Why these latencies exist? OS overhead with buffering, TCP time to queue messages, etc.
P receives message at [t + min2, t + RTT - min1]
- Lower Bound explained: at minimum, min1 could be 0, but we must incur latency of at least min2 → add to t to get time on P’s end
- Upper Bound explained: so min2 can be represented as RTT - min1 → add to t to get time on P’s end
Error bounded at most (RTT - min2 - min1) / 2
If min1 and min2 are unknown → set to 0