1
0
mirror of https://github.com/bitcoinbook/bitcoinbook synced 2024-11-27 02:18:25 +00:00

Merge branch 'chapter/Chapter8' into develop

This commit is contained in:
Andreas M. Antonopoulos 2014-08-07 13:38:52 -05:00
commit 1e4501283e

View File

@ -1,7 +1,5 @@
[[ch8]]
== Chapter 8 - Consensus & Mining
*DRAFT - DO NOT SUBMIT ISSUES OR PULL REQUESTS YET PLEASE - CONSTANT CHANGES HAPPENING*
== Mining and Consensus
[[mining]]
=== Introduction - Mining and Consensus
@ -553,12 +551,6 @@ As we saw above the target determines the difficulty and therefore affects how l
Bitcoin's blocks are generated every 10 minutes, on average. This is bitcoin's heartbeat and underpins the frequency of currency issuance and the speed of transaction settlement. It has to remain constant not just over the short term, but over a period of many decades. Over this time, it is expected that computer power will continue to increase at a rapid pace. Furthermore, the number of participants in mining and the computers they use will also constantly change. To keep the block generation time at 10 minutes, the difficulty of mining must be adjusted to account for these changes. In fact, difficulty is a dynamic parameter that will be periodically adjusted to meet a 10-minute block target. In simple terms, the difficulty target is set to whatever mining power will result in a 10-minute block interval.
In the chart below, we see the bitcoin network's hashing power increase over the past two years. As you can see, the competition between miners and the growth of bitcoin has resulted in an exponential increase in the hashing power (total hashes per second across the network):
[[network_hashing_power]]
.Total hashing power, giga-hashes per second, over two years
image::images/NetworkHashingRate.png["NetworkHashingRate"]
How then is such an adjustment made in a completely de-centralized network? Difficulty re-targeting occurs automatically and on every full node independently. Every 2016 blocks, all nodes re-target the Proof-of-Work difficulty. The equation for retargeting difficulty measures the time it took to find the last 2016 blocks and compares that to the expected time of 20160 minutes (two weeks based upon a desired 10 minute block time). The ratio between the actual timespan and desired timespan is calculated and a corresponding adjustment (up or down) is made to the difficulty. In simple terms: If the network is finding blocks faster than every 10 minutes, the difficulty increases. If block discovery is slower than expected, the difficulty decreases.
The equation can be summarized as:
@ -610,12 +602,6 @@ To avoid extreme volatility in the difficulty, the retargeting adjustment must b
The difficulty of finding a bitcoin block is approximately '10 minutes of processing' for the entire network, based on the time it took to find the previous 2016 blocks, adjusted every 2016 blocks.
====
The periodic adjustment of difficulty to match the amount of hashing power available and to keep the block generation time constant at 10 minutes has resulted in an exponential increase in difficulty. As the amount of hashing power applied to mining bitcoin has exploded, the difficulty has risen to match it. The difficulty metric is measured as a ratio of existing difficulty over minimum difficulty (the difficulty of the first block):
[[bitcoin_difficulty]]
.Bitcoin's mining difficulty metric, over two years
image::images/BitcoinDifficulty.png["BitcoinDifficulty"]
Note that the target difficulty is independent of the number of transactions or the value of transactions. This means that the amount of hashing power and therefore electricity expended to secure bitcoin is also entirely independent of the number of transactions. Bitcoin can scale up, achieve broader adoption and remain secure without any increase in hashing power from today's level. The increase in hashing power represents market forces as new miners enter the market to compete for the reward. As long as enough hashing power is under the control of miners acting honestly in pursuit of the reward, it is enough to prevent "takeover" attacks and therefore it is enough to secure bitcoin.
The target difficulty is closely related to the cost of electricity and the exchange rate of bitcoin vis-a-vis the currency used to pay for electricity. High performance mining systems are about as efficient as possible with the current generation of silicon fabrication, converting electricity into hashing computation at the highest rate possible. The primary influence on the mining market is the price of one kilowatt-hour in bitcoin, as that determines the profitability of mining and therefore the incentives to enter or exit the mining market.
@ -643,7 +629,7 @@ When a node receives a new block, it will validate the block by checking it agai
* Check the first transaction (and only the first) is a coinbase generation transaction
* Validate all transactions within the block, using the transaction checklist discussed in <<tx_verification>>
The independent validation of each new block by every node on the network ensures that the miners can't cheat. In previous sections we saw how the miners get to write a transaction that awards them the new bitcoins created within the block and claim the transaction fees. Why does the miner not write themselves a transaction for a thousand bitcoin instead of the correct reward? Because that would make the block invalid, which would result in it being rejected and therefore that transaction would never become part of the ledger. The miner has to construct a perfect block, based on the shared rules that all nodes follow and mine it with a correct solution to the Proof-of-Work. To do so they expend a lot of electricity in mining and if they cheat all the electricity and effort is wasted. This is why independent validation is a key component of decentralized consensus.
The independent validation of each new block by every node on the network ensures that the miners can't cheat. In previous sections we saw how the miners get to write a transaction that awards them the new bitcoins created within the block and claim the transaction fees. Why doesn't the miner write themselves a transaction for a thousand bitcoin instead of the correct reward? Because every node validates blocks according to the same rules. An invalid coinbase transaction would make the entire block invalid, which would result in the block being rejected and therefore that transaction would never become part of the ledger. The miner has to construct a perfect block, based on the shared rules that all nodes follow and mine it with a correct solution to the Proof-of-Work. To do so they expend a lot of electricity in mining and if they cheat all the electricity and effort is wasted. This is why independent validation is a key component of decentralized consensus.
=== Assembling and Selecting Chains of Blocks
@ -655,18 +641,18 @@ The "main chain" at any time is whichever chain of blocks has the most cumulativ
When a new block is received, a node will try to slot it into the existing blockchain. The node will look at the block's "previous block hash" field, which is the reference to the new block's parent. Then the node will attempt to find that parent in the existing blockchain. Most of the time, the parent will be the "tip" of the main chain, meaning this new block extends the main chain. For example, the new block 277,316 has a reference to the hash of its parent block 277,315. Most nodes that receive 277,316 will already have block 277,315 as the tip of their main chain and will therefore link the new block and extend that chain.
Sometimes, as we will see in <<forks>>, the new block extends a chain that is not the main chain. In that case, the node will attach the new block to the secondary chain it extends and then compare the difficulty of the secondary chain to the main chain. If the secondary chain has more cumulative difficulty than the main chain, the node will _reconverge_ on the secondary chain, meaning it will select the secondary chain as its new main chain, making the old main chain a secondary chain.
Sometimes, as we will see in <<forks>>, the new block extends a chain that is not the main chain. In that case, the node will attach the new block to the secondary chain it extends and then compare the difficulty of the secondary chain to the main chain. If the secondary chain has more cumulative difficulty than the main chain, the node will _reconverge_ on the secondary chain, meaning it will select the secondary chain as its new main chain, making the old main chain a secondary chain. If the node is a miner, it will now construct a block extending this new, longer, chain.
If a block is received and no parent is found in the existing chains, then that block is considered an "orphan". Orphan blocks are put into a temporary pool where they will stay until their parent is received. Once the parent is received and linked into the existing chains, the orphan can be pulled out of the orphan pool and linked to the parent, making it part of a chain. Orphan blocks usually occur when two blocks that were mined within a short time of each other are received in reverse order (child before parent).
If a valid block is received and no parent is found in the existing chains, then that block is considered an "orphan". Orphan blocks are saved in the orphan block pool where they will stay until their parent is received. Once the parent is received and linked into the existing chains, the orphan can be pulled out of the orphan pool and linked to the parent, making it part of a chain. Orphan blocks usually occur when two blocks that were mined within a short time of each other are received in reverse order (child before parent).
By selecting the greatest-difficulty chain, all nodes eventually achieve network-wide consensus. Temporary discrepancies between chains are resolved eventually as more Proof-of-Work is added, extending one of the possible chains. Mining nodes "vote" with their mining power by choosing which chain to extend by mining the next block. When they find a new block and extend the chain, the new block itself represents their vote.
By selecting the greatest-difficulty chain, all nodes eventually achieve network-wide consensus. Temporary discrepancies between chains are resolved eventually as more Proof-of-Work is added, extending one of the possible chains. Mining nodes "vote" with their mining power by choosing which chain to extend by mining the next block. When they mine a new block and extend the chain, the new block itself represents their vote.
In the next section we will look at how discrepancies between competing chains (forks) are resolved by the independent selection of the longest difficulty chain.
[[forks]]
==== Blockchain Forks
Because the blockchain is a decentralized data structure, different copies of it are not always consistent. Blocks may arrive at different nodes at different times, causing them to have a different perspective of the blockchain. To resolve this, each node always selects and attempts to extend the chain of blocks that represents the most Proof-of-Work, also known as the longest chain or greatest cumulative difficulty chain. By summing the difficulty recorded in each block in a chain, a node can calculate the total amount of Proof-of-Work that has been expended to create that chain. As long as all nodes select the longest cumulative difficulty chain, the global bitcoin network eventually converges to a consistent state. Forks occur as temporary inconsistencies between versions of the blockchain, which are resolved by eventual re-convergence as more blocks are added to one of the forks.
Because the blockchain is a decentralized data structure, different copies of it are not always consistent. Blocks may arrive at different nodes at different times, causing the nodes to have different perspectives of the blockchain. To resolve this, each node always selects and attempts to extend the chain of blocks that represents the most Proof-of-Work, also known as the longest chain or greatest cumulative difficulty chain. By summing the difficulty recorded in each block in a chain, a node can calculate the total amount of Proof-of-Work that has been expended to create that chain. As long as all nodes select the longest cumulative difficulty chain, the global bitcoin network eventually converges to a consistent state. Forks occur as temporary inconsistencies between versions of the blockchain, which are resolved by eventual re-convergence as more blocks are added to one of the forks.
In the next few diagrams, we follow the progress of a "fork" event across the network. The diagram is a simplified representation of bitcoin as a global network. In reality, the bitcoin network's topology is not organized geographically. Rather, it forms a mesh network of interconnected nodes, which may be located very far from each other geographically. The representation of a geographic topology is a simplification used for the purposes of illustrating a fork. In the real bitcoin network, the "distance" between nodes is measured in "hops" from node to node, not in terms of their location. For illustration purposes, different blocks are shown as different colors, spreading across the network and coloring the connections they traverse.
@ -710,12 +696,90 @@ It is theoretically possible for a fork to extend to two blocks, if two blocks a
Bitcoin's block interval of 10 minutes is a design compromise between fast confirmation times (settlement of transactions) and the probability of a fork. A faster block time would make transactions clear faster but lead to more frequent blockchain forks, whereas a slower block time would decrease the number of forks but make settlement slower.
=== Mining Pools
{miners that are on mining pools get the difficulty (do not calculate difficulty independently) they are given the difficulty from the mining pool so they don't have to calculate the difficulty themselves and they are actually given a lower difficulty target. There are essentially two classifications of miners today - pool miners and solo miners. Solo miners run a full node and compete on their own. Whereas pool miners collaborate with one another and compete against the network as a team, while sharing the reward. The reason miners join pools - solo miners need an enormous amount of hashing power in order to have even the slimmest chance of finding a solution to a block which will make their earnings erratic. By participating in a pool, miners get smaller shares but a more regular share of rewards, reducing uncertainty. Solo mining is becoming obsolete, as the difficulty increases the likelihood of a solo miner finding a solution is more like winning the lottery.}
==== Managed Pools
==== P2Pool
=== Mining and the Hashing Race
Bitcoin mining is an extremely competitive industry. The hashing power has increased exponentially, every year of bitcoin's existence. Some years the growth has reflected a complete change of technology, such as in 2010 and 2011 when many miners switched from using CPU mining to Graphical Processing Unit (GPU) mining and Field Programmable Gate Array (FPGA) mining. In 2013 the introduction of Application Specific Integrated Circuit (ASIC) mining lead to another giant leap in mining power, by placing the SHA-256 function directly on silicon chips specialized for the purpose of mining. The first such chips could deliver more mining power in a single box than the entire bitcoin network in 2010.
* 2009 - 0.5 MH/sec to 8 MH/sec (x 16 growth)
* 2010 - 8 MH/sec to 116 GH/sec (x 14,500 growth)
* 2011 - 16 GH/sec - 9 TH/sec (x 562 growth)
* 2012 - 9 TH/sec - 23 TH/sec (x 2.5 growth)
* 2013 - 23 TH/sec - 10 PH/sec (x 450 growth)
* 2014 - 10 PH/sec - 150 PH/sec in August (x 15 growth)
In the chart below, we see the bitcoin network's hashing power increase over the past two years. As you can see, the competition between miners and the growth of bitcoin has resulted in an exponential increase in the hashing power (total hashes per second across the network):
[[network_hashing_power]]
.Total hashing power, giga-hashes per second, over two years
image::images/NetworkHashingRate.png["NetworkHashingRate"]
As the amount of hashing power applied to mining bitcoin has exploded, the difficulty has risen to match it. The difficulty metric in the following chart is measured as a ratio of current difficulty over minimum difficulty (the difficulty of the first block):
[[bitcoin_difficulty]]
.Bitcoin's mining difficulty metric, over two years
image::images/BitcoinDifficulty.png["BitcoinDifficulty"]
In the last two years, the ASIC mining chips have become denser and denser, approaching the cutting edge of silicon fabrication with a feature size (resolution) of 22 nanometers (nm). Currently, ASIC manufacturers are aiming to overtake general purpose CPU chip manufacturers, designing chips with a feature size of 16nm, because the profitability of mining is driving this industry even faster than general computing. There are no more giant leaps left in bitcoin mining, because the industry has reached the forefront of "Moore's Law". Still, the mining power of the network continues to advance at an exponential pace as the race for higher density chips is matched with a race for higher density data centers where thousands of these chips can be deployed. It's no longer about how much mining can be done with one chip but how many chips can be squeezed into a building, while still dissipating the heat and providing adequate power.
==== The Extra Nonce Solution
Since 2012 bitcoin mining has evolved to resolve a fundamental limitation in the structure of the block header. In the early days of bitcoin, a miner could find a block by iterating through the nonce until the resulting hash was below the target. As difficulty increased, miners often cycled through all 4 billion values of the nonce without finding a block. However, this was easily resolved by updating the block timestamp to account for the elapsed time. Since the timestamp is part of the header, the change would allow miners to iterate through the values of the nonce again with different results. Once mining hardware exceeded 4 GH/sec however, this approach became increasingly difficult as the nonce values were exhausted in less than a second. As ASIC mining equipment started pushing and then exceeding the TH/sec hash rate, the mining software needed more space for nonce values in order to find valid blocks. The timestamp could be stretched a bit, but moving it too far into the future would cause the block to become invalid. A new source of "change" was needed in the block header. The solution was to use the coinbase transaction as a source of extra nonce values. Since the coinbase script can store between 2 and 100 bytes of data, miners started using that space as extra nonce space, allowing them to explore a much larger range of block header values to find valid blocks. The coinbase transaction is included in the merkle tree, which means that any change in the coinbase script causes the merkle root to change. Eight bytes of extra nonce, plus the 4 bytes of "standard" nonce allow miners to explore a total 2^96^ (8 followed by 28 zeroes) possibilities *per second* without having to modify the timestamp. If, in the future, a miner could run through all these possibilities, they could then modify the timestamp. There is also more space in the coinbase script for future expansion of the extra nonce space.
==== Mining Pools
In this highly competitive environment, individual miners working alone (also known as solo miners) don't stand a chance. The likelihood of them finding a block to offset their electricity and hardware costs is so low that it represents a gamble, like playing the lottery. Even the fastest consumer ASIC mining system cannot keep up with commercial systems that stack tens of thousands of these chips in giant warehouses near hydro-electric power stations. Miners now collaborate to form mining pools, pooling their hashing power and sharing the reward among thousands of participants. By participating in a pool, miners get a smaller share of the overall reward, but typically get rewarded every day, reducing uncertainty.
Let's look at a specific example. Assume a miner has purchased mining hardware with a combined hashing rate of 6,000 giga-hashes per second (GH/s) or 6 TH/s. In August of 2014 this equipment costs approximately $10,000 USD. The hardware also consumes 3 kilowatts (kW) of electricity when running, 72 kW-hours a day, at a cost of $7 or $8 per day on average. At current bitcoin difficulty, the miner will be able to solo-mine a block approximately once every 155 days, or every 5 months. If the miner does find a single block in that timeframe, the payout of 25 bitcoin, at approximately $600 per bitcoin will result in a single payout of $15,000 which will cover the entire cost of the hardware and the electricity consumed over the time period, leaving a net profit of approximately $3,000. However, the chance of finding a block in a 5-month period depends on the miner's luck. They might find two blocks in 5 months and make a very large profit. Or they might not find a block for 10 months and suffer a financial loss. Even worse, the difficulty of the bitcoin proof-of-work algorithm is likely to go up significantly over that period, at the current rate of growth of hashing power, meaning the miner has at most 6 months to break-even before the hardware is effectively obsolete and must be replaced by more powerful mining hardware. If this miner participates in a mining pool, instead of waiting for a once-in-5-month $15,000 windfall, they will be able to earn approximately $500 to $750 per week. The regular payouts from a mining pool will help them amortize the cost of hardware and electricity over time without taking an enormous risk. The hardware will still be obsolete in 6-9 months and the risk is still high, but the revenue is at least regular and reliable over that period.
Mining pools coordinate many hundreds or thousands of miners, over specialized pool mining protocols. The individual miners configure their mining equipment to connect to a pool server, after creating an account with the pool. Their mining hardware remains connected to the pool server while mining, synchronizing their efforts with the other miners. Thus, the pool miners share the effort to mine a block and then share in the rewards.
Successful blocks pay the reward to a pool bitcoin address, rather than individual miners. The pool server will periodically make payments to the miner bitcoin addresses, once their share of the rewards has reached a certain threshold. Typically, the pool server charges a percentage fee of the rewards for providing the pool mining service.
Miners participating in a pool, split the work of searching for a solution to a candidate block, earning "shares" for their mining contribution. The mining pool sets a lower difficulty target for earning a share, typically more than 1,000 times easier than the bitcoin network's difficulty. When someone in the pool successfully mines a block, the reward is earned by the pool and then shared with all miners in proportion to the number of shares they contributed to the effort.
Pools are open to any miner, big or small, professional or amateur. A pool will therefore have some participants with a single small mining machine, others with a garage-full of high-end mining hardware. Some will be mining with a few tens of a kilowatt of electricity, others will be running a data center consuming a megawatt of power. How does a mining pool measure the individual contributions, so as to fairly distribute the rewards, without the possibility of cheating? The answer is to use bitcoin's Proof-of-Work algorithm to measure each pool miner's contribution, but set at a lower difficulty so that even the smallest pool miners win a share frequently enough to make it worthwhile to contribute to the pool. By setting a lower difficulty for earning shares, the pool measures the amount of work done by each miner. Each time a pool miner finds a block header hash that is less than the pool difficulty, they prove they have done the hashing work to find that result. More importantly, the work to find shares contributes, in a statistically measurable way, to the overall effort to find a hash lower than the bitcoin network's target. Thousands of miners trying to find low-value hashes will eventually find one low enough to satisfy the bitcoin network target.
Let's return to the analogy of a dice game. If the dice players are throwing dice with a goal of throwing less than four (the overall network difficulty), a pool would set an easier target, counting how many times the pool players managed to throw less than eight. When a pool player throws eight or less (the pool share target), they earn shares, but they don't win the game because they don't achieve the game target (less than four). The pool players will achieve the easier pool target much more often, earning them shares very regularly, even when they don't achieve the harder target of winning the game. Every now and then, one of the pool players will throw a combined dice throw of four or less and the pool wins. Then, the earnings can be distributed to the pool players based on the shares they earned. Even though the target of eight-or-less wasn't winning, it was a fair way to measure dice throws for the players and occasionally produces a less-than-four throw.
Similarly, a mining pool will set a pool difficulty that will ensure that an individual pool miner can find block header hashes that are less than the pool difficulty quite often, earning shares. Every now and then, one of these attempts will produce a block header hash that is less than the bitcoin network target, making it a valid block and the whole pool wins.
===== Managed Pools
Most mining pools are "managed", meaning that there is a company or individual running a pool server. The owner of the pool server is called the _pool operator_ and they charge pool miners a percentage fee of the earnings.
The pool server runs specialized software and a pool-mining protocol that coordinates the activities of the pool miners. The pool server is also connected to one or more full bitcoin nodes and has direct access to a full copy of the blockchain database. This allows the pool server to validate blocks and transactions on behalf of the pool miners, relieving them of the burden of running a full node. For pool miners, this is an important consideration, as a full node requires a dedicated computer with at least 15-20 gigabytes of persistent storage (disk) and at least 2 gigabytes of memory (RAM). Furthermore, the bitcoin software running on the full node needs to be monitored, maintained and upgraded frequently. Any downtime caused by a lack of maintenance or lack of resources will impact the miner's profitability. For many miners the ability to mine without running a full node is another big benefit of joining a managed pool.
Pool miners connect to the pool server using a mining protocol such as Stratum (STM) or GetBlockTemplate (GBT). An older standard called GetWork (GWK) is now mostly obsolete since late 2012, as it does not easily support mining at hash rates above 4 GH/s. Both the STM and GBT protocols create block _templates_ that contain a template of a candidate block header. The pool server constructs a candidate block by aggregating transactions, adding a coinbase transaction (with extra nonce space), calculating the merkle root and linking to the previous block hash. The header of the candidate block is then sent to each of the pool miners as a template. Each pool miner then mines using the block template, at a lower difficulty than the bitcoin network difficulty and sends any successful results back to the pool server to earn shares.
===== P2Pool
Managed pools create the possibility of cheating by the pool operator, who might direct the pool effort to double-spend transactions or invalidate blocks (see <<51pct>>). Furthermore, centralized pool servers represent a single-point-of-failure. If the pool server is down or is attacked by Denial-of-Service, the pool miners cannot mine. In 2011, to resolve these issues of centralization, a new pooled mining method was proposed and implemented: P2Pool is a peer-to-peer mining pool, without a central operator.
P2Pool works by de-centralizing the functions of the pool server, implementing a parallel blockchain-like system called a _sharechain_. A sharechain is a blockchain running at a lower difficulty than the bitcoin blockchain. The sharechain allows pool miners to collaborate in a de-centralized pool, by mining shares on the sharechain at a rate of one share block every 30 seconds. Each of the blocks on the sharechain records a proportionate share reward for the pool miners who contribute work, carrying the shares forward from the previous share block. When one of the share blocks also achieves the difficulty target of the bitcoin network it is propagated and included on the bitcoin blockchain, rewarding all the pool miners who contributed to the all the shares that preceded the winning share block. Essentially, instead of a pool server keeping track of pool miner shares and rewards, the sharechain allows all pool miners to keep track of all shares using a de-centralized consensus mechanism like bitcoin's blockchain consensus mechanism.
P2Pool mining is more complex than pool mining, as it requires that the pool miners run a dedicated computer with enough disk space, memory and internet bandwidth to support a full bitcoin node and the p2pool node software. P2Pool miners connect their mining hardware to their local p2pool node, which simulates the functions of a pool server by sending block templates to the mining hardware. On P2Pool, individual pool miners construct their own candidate blocks, aggregating transactions much like solo-miners but then mine collaboratively on the sharechain. P2Pool is a hybrid approach that has the advantage of much more granular payouts than solo mining, but without giving too much control to a pool operator like managed pools.
Recently, participation in P2Pool has increased significantly as mining concentration in mining pools has approached levels that create concerns of a 51% attack (see <<51pct>>). Further development of the P2Pool protocol continues with the expectation of removing the need for running a full node and therefore making de-centralized mining even easier to use.
[[51pct]]
[[consensus_attacks]]
=== Consensus Attacks
==== 51% Attack
==== Selfish Mining Attack
Bitcoin's consensus mechanism is, at least theoretically, vulnerable to attack by miners (or pools) that attempt to use their hashing power to dishonest or destructive ends. As we saw, the consensus mechanism depends on having a majority of the miners acting honestly out of self-interest. However, if a miner or group of miners can achieve a significant share of the mining power, they can attack the consensus mechanism so as to disrupt the security and availability of the bitcoin network.
It is important to note that consensus attacks can only affect future consensus, or at best the most recent past (tens of blocks). Bitcoin's ledger becomes more and more immutable as time passes. Beyond a certain "depth", blocks are absolutely immutable, even under a sustained consensus attack that causes a fork. Consensus attacks also do not affect the security of the private keys and signing algorithm (ECDSA). A consensus attack cannot steal bitcoins, spend bitcoins without signatures, redirect bitcoins or otherwise change past transactions or ownership records. Consensus attacks can only affect the most recent blocks and cause denial-of-service disruptions on the creation of future blocks.
One attack scenario against the consensus mechanism is called the "51% attack". In this scenario a group of miners, controlling a majority (51%) of the total network's hashing power, collude to attack bitcoin. With the ability to mine the majority of the blocks, the attacking miners can cause deliberate "forks" in the blockchain and double-spend transactions or execute denial-of-service attacks against specific transactions or addresses. A fork/double-spend attack is one where the attacker causes previously confirmed blocks to be invalidated by forking below them and re-converging on an alternate chain. With sufficient power, an attacker can invalidate six or more blocks in a row, causing transactions that were considered immutable (6 confirmations) to be invalidated. Note that a double-spend can only be done on the attacker's own transactions, for which the attacker can produce a valid signature. Double-spending one's own transactions is profitable if by invalidating a transaction the attacker can get a non-reversible exchange payment or product without paying for it.
In addition to a double-spend attack, the other scenario for a consensus attack is to deny service to specific bitcoin participants (specific bitcoin addresses). An attacker with a majority of the mining power can simply ignore specific transactions. If they are included in a block mined by another miner the attacker can deliberately fork and re-mine that block, again excluding the specific transactions. This type of attack can result in a sustained denial of service against a specific address or set of addresses for as long as the attacker controls the majority of the mining power.
Despite its name, the 51% attack scenario doesn't actually require 51% of the hashing power. In fact, such an attack can be attempted with a smaller percentage of the hashing power. The 51% threshold is simply the level at which such an attack is almost guaranteed to succeed. A consensus attack is essentially a tug-of-war for the next block and the "stronger" group is more likely to win. With less hashing power, the probability of success is reduced, as other miners control the generation of some blocks with their "honest" mining power. One way to look at it is that the more hashing power an attacker has, the longer the fork they can deliberately create, the more blocks in the recent past they can invalidate, or the more blocks in the future they can control. Security research groups have used statistical modeling to claim that various types of consensus attacks are possible with as little as 30% of the hashing power.
The massive increase of total hashing power has arguably made bitcoin impervious to attacks by a single miner. There is no possible way for a solo miner to control even 1% of the total mining power. However, the centralization of control caused by mining pools has introduced the risk of for-profit attack by a mining pool operator. The pool operator in a managed pool controls the construction of candidate blocks and also controls which transactions are included. This gives the pool operator the power to exclude transactions or introduce double-spend transactions. If such abuse of power is done in a limited and subtle way, a pool operator could conceivably profit from a consensus attack without being noticed.
Not all attackers will be motivated by profit, however. One potential attack scenario is where an attacker intends to disrupt the bitcoin network without the possibility of profiting from such disruption. A malicious attack aimed at crippling bitcoin would require enormous investment and covert planning, but could conceivably be launched by a well funded, most likely state-sponsored attacker. Alternatively, a well-funded attacker could attack bitcoin's consensus by simultaneously amassing mining hardware, compromising pool operators and attacking other pools with denial-of-service. All of these scenarios are theoretically possible, but increasingly impractical as the bitcoin network's overall hashing power continues to grow exponentially. Recent advancements in bitcoin, such as P2Pool mining, aim to further de-centralize mining control, making bitcoin consensus even harder to attack.
Undoubtedly, a serious consensus attack would erode confidence in bitcoin in the short term, possibly causing a significant price decline. However, the bitcoin network and software is constantly evolving, so consensus attacks would be met with immediate counter-measures by the bitcoin community, making bitcoin hardier, stealthier and more robust.