Bitcoin is structured as a peer-to-peer network architecture on top of the Internet. The term peer-to-peer or P2P means that the computers that participate in the network are peers to each other, they are all equal there are no "special" nodes and all nodes share the burden of providing network services. The network nodes interconnect in a mesh network with a "flat" topology. There is no "server", no centralized service and no hierarchy within the network. Nodes in a peer-to-peer network both provide and consume services at the same time, with reciprocity acting as the incentive for participation. Peer-to-peer networks are inherently resilient, de-centralized and open. The pre-eminent example of a P2P network architecture was the early Internet itself, where nodes on the IP network were equal. Today's Internet architecture is more hierarchical, but the Internet Protocol still retains its flat-topology essence. Beyond bitcoin, the largest and most successful application of P2P technologies is file sharing, with Napster as the pioneer and bittorrent as the most recent evolution of the architecture.
Bitcoins P2P network architecture is much more than a topology choice. Bitcoin is a peer-to-peer digital cash system by design, and the network architecture is both a reflection and a foundation of that core characteristic. De-centralization of control is a core design principle and that can only be achieved and maintained by a flat, de-centralized P2P consensus network.
The term "bitcoin network" refers to the collection of nodes running the bitcoin P2P protocol. In addition to the bitcoin P2P protocol, there are other protocols such as Stratum, that are used for mining and lightweight or mobile wallets. These additional protocols are provided by gateway routing servers that access the bitcoin network using the bitcoin P2P protocol and then extend that network to nodes running other protocols. For example, Stratum servers connect Stratum mining nodes via the Stratum protocol to the main bitcoin network and bridge the Stratum protocol to the bitcoin P2P protocol. We use the term "extended bitcoin network" to refer to the overall network that includes the bitcoin P2P protocol, pool mining protocols, the Stratum protocol and any other related protocols connecting the components of the bitcoin system.
While nodes in the bitcoin P2P network are equal, they may take on different "roles", depending on the functionality they are supporting. A bitcoin node is a collection of functions: routing, the blockchain database, mining, and wallet services. A full node with all four of these functions is shown below:
[[full_node_reference]]
.A bitcoin network node with all four functions: Network routing, Blockchain database, Mining and Wallet
All nodes include the routing function to participate in the network and may include other functionality. All nodes validate and propagate transactions and blocks, discover and maintain connections to peers. In the full node example above, the routing function is indicated by an orange circle named "Network Routing Node".
Some nodes, called full nodes, also maintain a complete and up-to-date copy of the blockchain. Full nodes can autonomously and authoritatively verify any transaction without external reference. Some nodes maintain only a subset of the blockchain and verify transactions using a method called _Simple Payment Verification_ or SPV. These nodes are known as SPV or Lightweight nodes. In the full node example above, the full node blockchain database function is indicated by a blue circle named "Blockchain Database". SPV nodes are drawn without the blue circle, showing that they do not have a full copy of the blockchain.
Mining nodes compete to create new blocks by running specialized hardware to solve the proof-of-work algorithm. Some mining nodes are also full nodes, maintaining a full copy of the blockchain while others are lightweight nodes participating in pool mining and depending on a pool server to maintain a full node. The mining function is shown in the full node above as a black circle named "Mining".
User wallets may be part of a full node, as is usually the case with desktop bitcoin clients. Increasingly many user wallets, especially those running on resource constrained devices such as smart phones, are SPV nodes. The wallet function is shown above as a green circle named "Wallet".
In addition to the main node types on the bitcoin P2P protocol, there are servers and nodes running other protocols, such as specialized mining pool protocols and lightweight client access protocols.
Here are the most common node types on the extended bitcoin network:
[[node_type_ledgend]]
.Different types of nodes on the extended bitcoin network
The main bitcoin network, running the bitcoin P2P protocol consists of between 7,000 to 10,000 nodes running various versions of the bitcoin reference client (Bitcoin Core) and a few hundred nodes running various other implementations of the bitcoin P2P protocol, such as BitcoinJ, Libbitcoin and btcd. A small percentage of the nodes on the bitcoin P2P network are also mining nodes, competing in the mining process, validating transactions and creating new blocks. Various large companies interface with the bitcoin network by running full-node clients based on the Bitcoin Core client, with full copies of the blockchain and a network node, but without mining or wallet functions. These nodes act as network edge routers, allowing various other services (exchanges, wallets, block explorers, merchant payment processing) to be built on top.
The extended bitcoin network includes the network running the bitcoin P2P protocol, described above, as well as nodes running specialized protocols. Attached to the main bitcoin P2P network are a number of pool servers and protocol gateways that connect nodes running other protocols, mostly pool mining nodes (see <<mining>>) and lightweight wallet clients, which do not carry a full copy of the blockchain.
The diagram below shows the extended bitcoin network with the various types of nodes, gateway servers, edge routers and wallet clients and the various protocols they use to connect to each other.
[[bitcoin_network]]
.The extended bitcoin network showing various node types, gateways and protocols
When a new node boots up, it must discover other bitcoin nodes on the network in order to participate. To start this process, a new node must discover at least one existing node on the network and connect to it. The geographic location of the other nodes is irrelevant, the bitcoin network topology is not geographically defined. Therefore, any existing bitcoin nodes can be selected at random.
To connect to a known peer, nodes establish a TCP connection, usually to port 8333 (the bitcoin "well known" port), or an alternative port if one is provided. Upon establishing a connection, the node will start a "handshake" by transmitting a +version+ message, which contains basic identifying information, including:
(See https://github.com/bitcoin/bitcoin/blob/d3cb2b8acfce36d359262b4afd7e7235eff106b0/src/net.cpp#L562 for an example of the +version+ network message)
The peer node responds with +verack+ to acknowledge and establish a connection, and optionally sends its own +version+ message if it wishes to reciprocate the connection and connect back as a peer.
How does a new node find peers? While there are no special nodes in bitcoin, there are some long running stable nodes that are listed in the client as _seed nodes_. While a new node does not have to connect with the seed nodes, it can use them to quickly discover other nodes in the network. In the Bitcoin Core client, the option to use the seed nodes is controlled by the option switch +-dnsseed+, which is set to 1, to use the seed nodes, by default. Alternatively, a bootstrapping node that knows nothing of the network must be given the IP address of at least one bitcoin node after which it can establish connections through further introductions. The command line argument +-seednode+ can be used to connect to one node just for introductions, using it as a DNS seed. After the initial seed node is used to form introductions, the client will disconnect from it and use the newly discovered peers.
Once one or more connections is established, the new node will send an +addr+ message containing its own IP address, to its neighbors. The neighbors will in turn forward the +addr+ message to their neighbors, ensuring that the newly connected node becomes well known and better connected. Additionally, the newly connected node can send +getaddr+ to the neighbors asking them to return a list of IP addresses of other peers. That way, a node can find peers to connect to and advertise its existence on the network for other nodes to find it. On a node running the Bitcoin Core client, you can list the peer connections with the command +getpeerinfo+:
A node must connect to a few different peers in order to establish diverse paths into the bitcoin network. These paths are not reliable, nodes come and go, and so the node must continue to discover new nodes as it loses old connections as well as assist other nodes when they bootstrap. Only one connection is needed to bootstrap, as the first node can offer introductions to its peer nodes and those peers can offer further introductions. Its also unnecessary and wasteful of network resources to connect to more than a handful of nodes. After bootstrapping a node will remember its most recent successful peer connections, so that if it is rebooted it can quickly reestablish connections with its former peer network. If none of the former peers respond to its connection request, the node can use the seed nodes to bootstrap again.
To override the automatic management of peers and to specify a list of IP addresses, users can provide the option +-connect=<IPAddress>+ and specify one or more IP addresses. If this option is used, the node will only connect to the selected IP addresses, instead of discovering and maintaining the peer connections automatically.
If there is no traffic on a connection, nodes will periodically send a message to maintain the connection. If a node has not communicated on a connection for more than 90 minutes it is assumed to be disconnected and a new peer will be sought. Thus the network dynamically adjusts to transient nodes, network problems, and can organically grow and shrink as needed without any central control.
=== Full Nodes
Full nodes are nodes that maintain a full blockchain. More accurately they probably should be called "full blockchain nodes". In the early years of bitcoin, all nodes were full nodes and currently the Bitcoin Core client is a full blockchain node. In the last two years however, new forms of bitcoin clients have been introduced, which do not maintain a full blockchain but run as lightweight clients. These are examined in more detail in the next section.
Full blockchain nodes maintain a complete and up-to-date copy of the bitcoin blockchain, which they independently build and verify, starting with the very first block (genesis block) and up to the latest known block in the network. A full blockchain node can independently and authoritatively verify any transaction, without recourse or reliance on any other node or source of information. The full blockchain node relies on the network to receive updates about new blocks of transactions, which it then verifies and incorporates into its local copy of the blockchain.
Running a full blockchain node gives you the pure bitcoin experience: independent verification of all transactions without the need to rely on, or trust, any other systems. It's easy to tell if you're running a full node because it requires several gigabytes of persistent storage (disk space) to store the full blockchain. If you need a lot of disk and it takes 2-3 days to "sync" to the network you are running a full node. That is the price of complete independence and freedom from central authority.
There are a few alternative implementations of full-blockchain bitcoin clients, built using different programming languages and software architectures. However, the most common implementation is the reference client Bitcoin Core, also known as the Satoshi Client. More than 90% of the nodes on the bitcoin network run various versions of Bitcoin Core. It is identified as "Satoshi" in the sub-version string sent in the +version+ message and shown by the command +getpeerinfo+ as we saw above, for example +/Satoshi:0.8.6/+.
=== Exchanging "Inventory"
The first thing a full node will do once it connects to peers is try to construct a complete blockchain. If it is a brand-new node and has no blockchain at all, then it only knows one block (the genesis block), which is statically embedded in the client software. Starting with block #0, the genesis block, the new node will have to download hundreds of thousands of blocks to synchronize with the network and establish a full blockchain.
The process of "syncing" the blockchain starts with the +version+ message, as that contains +BestHeight+, a node's current blockchain height (number of blocks). A node will see the +version+ messages from its peers, know how many blocks they each have and be able to compare to how many blocks it has in its own blockchain. Peered nodes will exchange a +getblocks+ message that contains the hash (fingerprint) of the top block on their local blockchain. One of the peers will be able to identify the received hash as belonging to a block that is not at the top, but rather belongs to an older block, thus deducing that its own local blockchain is longer than its peer's.
The peer that has the longer blockchain has more blocks that the other node and can identify which blocks the other node needs to "catch up". It will identify the first 500 blocks to share and transmit their hashes using an +inv+ (inventory) message. The node missing these blocks will then retrieve them, by issuing a series of +getdata+ messages requesting the full block data and identifying the requested blocks using the hashes from the +inv+ message.
Let's assume for example that a node only has the genesis block. It will then receive an +inv+ message from its peers containing the hashes of the next 500 blocks in the chain. It will start requesting blocks from all its connected peers, spreading the load and ensuring that it doesn't overwhelm any peer with requests. The node keeps track of how many blocks are "in transit" per peer connection, meaning blocks that it has requested but not received, checking that it does not exceed a limit (MAX_BLOCKS_IN_TRANSIT_PER_PEER). This way, if it needs a lot of blocks, it will only request new ones as previous requests are fulfilled, allowing the peers to control the pace of updates and not overwhelming the network. As each block is received, it is added to the blockchain as we will see in the next chapter <<blockchain>>. The local blockchain is gradually built up, more blocks are requested and received and the process continues until the node catches up to the rest of the network.
This process of comparing the local blockchain with the peers and retrieving any missing blocks happens any time a node goes offline for any period of time. Whether a node has been offline for a few minutes and is missing a few blocks, or a month and is missing a few thousand blocks, it starts by sending +getblocks+, gets an +inv+ response and starts downloading the missing blocks.
In the previous chapter we saw how wallet software creates transactions by collecting UTXO, providing the appropriate unlocking scripts and then constructing new outputs assigned to a new owner. The resulting transaction is then sent to the neighboring nodes in the bitcoin network so that it may be propagated across the entire bitcoin network.
Every bitcoin node that receives a transaction will first verify the transaction before forwarding it to its neighbors. This ensures that only valid transactions are propagated across the network, while invalid transactions are discarded at the first node that encounters them.
* Check none of the inputs have hash=0, N=-1 (coinbase transactions should not be relayed)
* Check that nLockTime is less than or equal to INT_MAX
* Check that the transaction size in bytes is greater than or equal to 100
* Check the number of signature operations contained in the transaction is less than the signature operation limit
* Reject "nonstandard" transactions: unlocking script (scriptSig) doing anything other than pushing numbers on the stack, or the locking script (scriptPubkey) not matching isStandard forms
* For each input, look in the main branch and the transaction pool to find the referenced output transaction. If the output transaction is missing for any input, this will be an orphan transaction. Add to the orphan transactions, if a matching transaction is not already in the pool.
* For each input, if the referenced output transaction is a coinbase output, it must have at least COINBASE_MATURITY (100) confirmations; else reject this transaction
* For each input, if the referenced output does not exist (e.g. never existed or has already been spent), reject this transaction
* Using the referenced output transactions to get input values, check that each input value, as well as the sum, are in the allowed range of values (less than 21m coins, more than 0)
* Verify the unlocking scripts for each input against the corresponding output locking scripts
These conditions can be seen in detail in the functions AcceptToMemoryPool, CheckTransaction and CheckInputs in the bitcoin reference client. Note that the conditions change over time, to address new types of Denial-of-Service attacks or sometimes to relax the rules so as to include more types of transactions.
By independently verifying each transaction as it is received and before propagating it, every node builds a pool of valid new transactions (the transaction pool), roughly in the same order.
Almost every node on the bitcoin network maintains a temporary list of unconfirmed transactions called the memory pool or transaction pool. Once a transaction is verified using the detailed checklist introduced in the section above, it is added to the transaction pool. Nodes use this pool to keep track of transactions that are known to the network but are not yet included in the blockchain. For example, a node that holds a users wallet will use the transaction pool to track incoming payments to the users wallet that have been received on the network but are not yet confirmed. Every node also maintains a separate pool of orphaned transactions as detailed in <<orphan_transactions>>. If a transactions inputs refer to a transaction that is not yet known, a missing parent, then it will be stored temporarily in the orphan pool until the parent transaction arrives. Both the transaction pool and orphan pool are stored in local memory and are not saved on persistent storage, rather they are dynamically populated from incoming network messages. When a node starts, both pools are empty and are gradually populated with new transactions received on the network.
As transactions are received and verified using the criteria in the previous section, they are added to the transaction pool and relayed to the neighboring nodes to propagate on the network.
When a transaction is added to the transaction pool, the orphan pool is checked for any orphans that reference this transaction's outputs (its children). Any orphans found are pulled from the orphan pool and validated using the above checklist. If valid, they are also added to the transaction pool, completing the chain that started with the parent transaction. In light of the newly added transaction which is no longer an orphan, the process is repeated recursively looking for any further descendants, until no more descendants are found. Through this process, the arrival of a parent transaction triggers a cascade reconstruction of an entire chain of interdependent transactions by re-uniting the orphans with their parents all the way down the chain.
Some implementations of the bitcoin client also maintain a UTXO pool which is the set of all unspent outputs on the blockchain. This may be housed in local memory or as an indexed database table on persistent storage. Unlike the transaction and orphan pools, the UTXO pool is not initialized empty but instead contains millions of entries of unspent transaction outputs including some dating back to 2009. Whereas the transaction and orphan pools represent a single nodes local perspective and may vary significantly from node to node depending upon when the node was started or restarted, the UTXO pool represents the emergent consensus of the network and therefore will vary little between nodes. Furthermore the transaction and orphan pools only contain unconfirmed transactions, while the UTXO pool only contains confirmed outputs.
As part of populating the block header, a mining node will create a summary of all the transactions added to the block. This summary is created by computing the _root_ of the Merkle Tree, which is a binary hash tree data structure. The merkle root is a 32-byte hash that provides a shortcut to identify individual transactions contained within that block.
A _Merkle Tree_, also known as a _Binary Hash Tree_ is a data structure created by Ralph Merkle used for efficiently summarizing and verifying the integrity of large sets of data. Merkle Trees are binary trees containing cryptographic hashes. When N data elements are hashed and summarized in a Merkle Tree, you can check to see if any one data element is included in the tree with at most +2*log~2~(N)+ calculations, making this a very efficient data structure. The term "tree" is used in computer science to describe a branching data structure, but these trees are usually displayed upside down with the "root" at the top and the "leaves" at the bottom of a diagram, as you will see in the examples that follow.
Merkle trees are used in bitcoin to summarize all the transactions in a block, producing an overall digital fingerprint of the entire set of transactions, which can be used to prove that a transaction is included in the set. A merkle tree is constructed by recursively hashing pairs of nodes until there is only one hash, called the _root_, or _merkle root_. The cryptographic hash algorithm used in bitcoin's merkle trees is SHA256 applied twice, also known as double-SHA256.
The merkle tree is constructed bottom-up. In the example below, we start with four transactions A, B, C and D, which form the _leaves_ of the Merkle Tree, shown in the diagram at the bottom. The transactions are not stored in the merkle tree, rather their data is hashed and the resulting hash is stored in each leaf node as H~A~, H~B~, H~C~ and H~D~:
+H~A~ = SHA256(SHA256(Transaction A))+
Consecutive pairs of leaf nodes are then summarized in a parent node, by concatenating the two hashes and hashing them together. For example, to construct the parent node H~AB~, the two 32-byte hashes of the children are concatenated to create a 64-byte string. That string is then double-hashed to produce the parent node's hash:
+H~AB~ = SHA256(SHA256(H~A~ + H~B~))+
The process continues until there is only one node at the top, the node known as the Merkle Root. That 32-byte hash is stored in the block header and summarizes all the data in all four transactions.
[[simple_merkle]]
.Calculating the nodes in a Merkle Tree
image::images/MerkleTree.png["merkle_tree"]
Since the merkle tree is a binary tree, it needs an even number of leaf nodes. If there is an odd number of transactions to summarize, the last transaction hash is duplicated to create an even number of leaf nodes, also known as a _balanced tree_. This is shown in the example below, where transaction C is duplicated:
[[merkle_tree_odd]]
.An even number of data elements, by duplicating one data element
The same method for constructing a tree from four transactions can be generalized to construct trees of any size. In bitcoin it is common to have several hundred to more than a thousand transactions in a single block, which are summarized in exactly the same way producing just 32-bytes of data from a single merkle root. In the diagram below, you will see a tree built from 16 transactions:
To prove that a specific transaction is included in a block, a node need only produce +log~2~(N)+ 32-byte hashes, constituting an _authentication path_ or _merkle path_ connecting the specific transaction to the root of the tree. This is especially important as the number of transactions increases, because the base-2 logarithm of the number of transactions increases much more slowly. This allows bitcoin nodes to efficiently produce paths of ten or twelve hashes (320-384 bytes) which can provide proof of a single transaction out of more than a thousand transactions in a megabyte sized block. In the example below, a node can prove that a transaction K is included in the block by producing a merkle path that is only four 32-byte hashes long (128 bytes total). The path consists of the four hashes H~L~, H~IJ~, H~MNOP~ and H~ABCDEFGH~. With those four hashes provided as an authentication path, any node can prove that H~K~ is included in the merkle root by computing four additional pair-wise hashes H~KL~, H~IJKL~ and H~IJKLMNOP~ that lead to the merkle root.
[[merkle_tree_path]]
.A Merkle Path used to prove inclusion of a data element
As you can see from the table above, while the block size increases rapidly, from 4KB with 16 transactions to a block size of 16 MB to fit 65,535 transactions, the merkle path required to prove the inclusion of a transaction increases much more slowly, from 128 bytes to only 512 bytes. With merkle trees, a node can download just the block headers (80 bytes per block) and still be able to identify a transaction's inclusion in a block by retrieving a small merkle path from a full node, without storing or transmitting the vast majority of the blockchain which may be several gigabytes in size. Nodes which do not maintain a full blockchain, called Simple Payment Verification or SPV nodes use merkle paths to verify transactions without downloading full blocks.