The Sui network is operated by a set of independent validators, each running its own instance of the Sui software on a separate machine (or a sharded cluster of machines operated by the same entity). A validator participates in the network by handling read and write requests sent by clients. This section focuses on the latter.
Sui uses proof of stake (PoS) to determine which validators operate the network and their voting power. Validators are incentivized to participate in good faith via a share of transaction fees, staking rewards, and slashing to punish misbehavior.
Operation of the Sui network is temporally partitioned into non-overlapping, approximate fixed-duration (e.g. 24-hour) epochs. During a particular epoch, the set of validators participating in the network is fixed. At an epoch boundary, reconfiguration may occur and can change the set of validators participating in the network and their voting power. Conceptually, reconfiguration starts a new instance of the Sui protocol with the previous epoch's final state as genesis and the new set of validators as the operators.
A quorum is a set of validators whose combined voting power is >2/3 of the total during a particular epoch. For example, in a Sui instance operated by four validators that all have the same voting power, any group containing three validators is a quorum.
The quorum size of >2/3 is chosen to ensure Byzantine fault tolerance (BFT). As we will see, a validator will commit a transaction (i.e., durably store the transaction and update its internal state with the effects of the transaction) only if it is accompanied by cryptographic signatures from a quorum. We call the combination of the transaction and the quorum signatures on its bytes a certificate. The policy of committing only certificates ensures Byzantine fault tolerance: if >2/3 of the validators faithfully follow the protocol, they are guaranteed to eventually agree on both the set of committed certificates and their effects.
A validator can handle two types of write requests: transactions and certificates. At a high level, a client:
- communicates a transaction to a quorum of validators to collect the signatures required to form a certificate.
- submits a certificate to a validator to commit state changes on that validator.
When a validator receives a transaction from a client, it will first perform transaction validity checks (e.g., validity of the sender's signature). If the checks pass, the validator locks all owned-objects and signs the transaction bytes. It then returns the signature to the client. The client repeats this process with multiple validators until it has collected signatures on its transaction from a committee, thereby forming a certificate.
Note that the process of collecting validator signatures on a transaction into a certificate and the process of submitting certificates can be performed in parallel. The client can simultaneously multicast transactions/certificates to an arbitrary number of validators. Alternatively, a client can outsource either or both of these tasks to a third-party service provider. This provider must be trusted for liveness (e.g., it can refuse to form a certificate), but not for safety (e.g., it cannot change the effects of the transaction, and does not need the user's secret key).
Once the client forms a certificate, it submits the certificate to a validator, which will perform certificate validity checks (e.g., ensuring the signers are validators in the current epoch, and the signatures are cryptographically valid). If the checks pass, the authority will execute the transaction inside the certificate. Execution of a transaction will either succeed and commit all of its effects to the ledger, or abort (e.g., due to an explicit
abort instruction, a runtime error such as division by zero, or exceeding the maximum gas budget) and have no effects other than debiting the transaction's gas input. In either case, the transaction will durably store the certificate indexed by the hash of its inner transaction.
As with transactions, we note that the process of sharing a certificate with validators can be parallelized and (if desired) outsourced to a third-party service provider. A client should broadcast its certificate to >1/3 of the validators to ensure that (up to BFT assumptions) at least one honest validator has executed and committed the certificate. Other validators may learn about the certificate via inter-validator state sync or via client-assisted state sync.
The role of Narwhal and Tusk
Sui takes advantage of Narwhal and Tusk: A DAG-based Mempool and Efficient BFT Consensus. Narwhal/Tusk (N/T) are also being implemented by Mysten Labs so that when Byzantine agreement is required we use a high-throughput DAG-based consensus to manage shared locks while execution on different shared objects is parallelized.
Narwhal enables the parallel ordering of transactions into batches that are collected into concurrently proposed blocks, and Tusk defines an algorithm for executing the DAG that these blocks form. N/T combined builds a DAG of blocks, concurrently proposed, and creates an order between those blocks as a byproduct of the building of the DAG. But that order is overlaid on top of the causal order of Sui transactions (the "payload" of Narwhal/Tusk here), and does not substitute for it:
- Narwhal/Tusk operates in OX, rather than XO mode (O = order, X = execute); the execution occurs after the Narwhal/Tusk ordering.
- The output of N/T is therefore a sequence of transactions, with interdependencies stored in the transaction data itself.
Consensus sequences certificates of transactions. These represent transactions that have already been presented to 2/3 of validators that checked that all their owned objects are available to be operated on and signed the transaction. Upon a certificate being sequenced, what we do is set the lock of the shared objects at the next available version to map to the execution of that certificate. So for example if we have a shared object X at version 2, and we sequence certificate T, we store T -> [(X, 2)]. That is all we do when we reach consensus, and as a result we are able to ingest a lot of sequenced transactions.
Now, once this is done Sui can execute all certificates that had their locks set, on one or multiple cores (currently). Obviously, transactions for earlier versions of objects need to be processed first (causally), and that reduces the degree of concurrency. The read and write set of the transaction can be statically determined from its versioned object inputs--execution can only read/write an object that was an input to the transaction, or that was created by the transaction.