What is the simplest setting where randomization can help solve consensus? Assume lock-step (synchrony) with $f<n$ crash failures. We know that in the worst case reaching agreement takes at least $f+1$ rounds. This lower bound holds even if the protocol is randomized so the natural question is:
Can randomization help reduce the expected number of rounds?
This post can be used as a prequel for our posts on how to use randomness to solve asynchronous agreement.
Perfect common coin
Assume that in each round $j$, parties have access to a perfect common coin, coin_j
. The coin for round $j$ is unpredictable before the end of round $j$ and is uniformly distributed. A perfect coin is a strong assumption. We will show in later posts build it.
Binary agreement for crash failures with a perfect common coin
Each party has an input 0 or 1 and the goal is to output a common value (agreement) that is an input value (validity).
value := input
round j:
send <value> to all parties
if coin_j = value, then output value
if hear both <0> and <1>, then value := coin_j
This protocol is so simple: if the coin equals your value then decide. Otherwise change to the coin value if you hear both 0 and 1.
Adding a termination gadget is also easy:
Once you output value,
then send <decide value> and terminate
Once you see <decide value>,
then output value
Analysis
Validity: clearly if all parties start with $b$ then they will never change their value.
Agreement: consider the first round $v^\star$ where some party $i$ outputs $b$. So party $i$ did not crash before sending $b$ to all parties and the coin in round $v^\star$ is $b$. At the end of round $v^\star$ all parties heard the value $b$ from $i$ so will either use the coin to set their value to $b$ or keep the value $b$. Similar to the validity argument above, in any round $> v^\star$, since all parties start with $b$ then they will leave with $b$.
Expected Termination: Conditioned on no party outputting a value in any round $<j$, consider the first party to not crash in round $j$. With probability $1/2$ the coin will be equal to its values and it will output its value.
So we have shown that in expected 2 rounds there will be at least one party that outputs a value.
Conditioned on a party outputting value $b$ in round $<j’$, we know from agreement and validity that all parties enter round $j’$ with the value $b$. With probability $1/2$ the coin in round $j’$ will be $b$ and all non-faulty parties will output their value in round $j’$.
So we have shown that in expected 4 rounds all non-faulty parties output a value. So terminate in expected 5 rounds.
In other words, given a perfect common coin, for any $f<n$ the expected number of rounds till all non-faulty parties terminate is constant.
Zooming out
How did randomization help? we used the common coin as a virtual leader. To maintain validity, we only listen to the virtual leader when we hear both 0 and 1. For safety we decide only when the condition coin_j = value
is true, this guarantees all other parties assign coin_j = value
.
Finally, why does the protocol end in an expected 5 rounds? it’s because we assumed the coin is unpredictable! If the adversary knew the coin values in advance it could force a $f+1$ round execution.
By assuming the coin is unpredictable we implicitly assume there is a limit to the adaptive power of the adversary. It must send its round $j$ message before learning the round $j$ coin. In later posts, we will be even more explicit about the adaptive power of the adversary.
Exercise: consider a super adaptive adversary with future-vision that at the beginning of round $j$ knows in advance the coin value and can decide what parties to crash in round $j$ based on this information. Show a strategy for this adversary that forces a $f+1$ round execution.
The protocol relies heavily on assuming a perfect common coin. In the next post, we will solve consensus for omission failures without this assumption by building a weak common coin.
Acknowledgments
Many thanks to Sravya Yandamuri and Naama Ben-David for insightful discussions and comments.
Your thoughts (and solution to the exercise) on Twitter.