bitpro가 원문에서 글자만 모은 내용이에요. 표·그림·수식은 빠지거나 깨질 수 있어요. 전체 3쪽 중 1쪽
Avalanche Platform
2020/06/30
Kevin Sekniqi, Daniel Laine, Stephen Buttolph, and Emin G¨ un Sirer
Abstract. This paper provides an architectural overview of the first release of theAvalanche platform,
codenamed Avalanche Borealis . For details on the economics of the native token, labeled $AVAX, we5
guide the reader to the accompanying token dynamics paper [2].
Disclosure: The information described in this paper is preliminary and subject to change at any time.
Furthermore, this paper may contain “forward-looking statements.” 1
Git Commit: 7497e4a4ba0a1ea2dc2a111bc6deefbf3023708e
1 Introduction10
This paper provides an architectural overview of the Avalanche platform. The key focus is on the three key
differentiators of the platform: the engine, the architectural model, and the governance mechanism.
1.1 Avalanche Goals and Principles
Avalanche is a high-performance, scalable, customizable, and secure blockchain platform. It targets three
broad use cases:15
– Building application-specific blockchains, spanning permissioned (private) and permissionless (public)
deployments.
– Building and launching highly scalable and decentralized applications (Dapps).
– Building arbitrarily complex digital assets with custom rules, covenants, and riders (smart assets).
1 Forward-looking statements generally relate to future events or our future performance. This includes, but is not
limited to, Avalanche’s projected performance; the expected development of its business and projects; execution
of its vision and growth strategy; and completion of projects that are currently underway, in development or
otherwise under consideration. Forward-looking statements represent our management’s beliefs and assumptions
only as of the date of this presentation. These statements are not guarantees of future performance and undue
reliance should not be placed on them. Such forward-looking statements necessarily involve known and unknown
risks, which may cause actual performance and results in future periods to differ materially from any projections
expressed or implied herein. Avalanche undertakes no obligation to update forward-looking statements. Although
forward-looking statements are our best prediction at the time they are made, there can be no assurance that they
will prove to be accurate, as actual results and future events could differ materially. The reader is cautioned not
to place undue reliance on forward-looking statements.
2 Kevin Sekniqi, Daniel Laine, Stephen Buttolph, and Emin G¨ un Sirer
The overarching aim of Avalanche is to provide a unifying platform for the creation, transfer, and trade of20
digital assets.
By construction, Avalanche possesses the following properties:
Scalable Avalanche is designed to be massively scalable, robust, and efficient. The core consensus engine
is able to support a global network of potentially hundreds of millions of internet-connected, low and high-
powered devices that operate seamlessly, with low latencies and very high transactions per second.25
Secure Avalanche is designed to be robust and achieve high security. Classical consensus protocols are
designed to withstand up to f attackers, and fail completely when faced with an attacker of size f + 1 or
larger, and Nakamoto consensus provides no security when 51% of the miners are Byzantine. In contrast,
Avalanche provides a very strong guarantee of safety when the attacker is below a certain threshold, which
can be parametrized by the system designer, and it provides graceful degradation when the attacker exceeds30
this threshold. It can uphold safety (but not liveness) guarantees even when the attacker exceeds 51%. It is
the first permissionless system to provide such strong security guarantees.
Decentralized Avalanche is designed to provide unprecedented decentralization. This implies a commitment
to multiple client implementations and no centralized control of any kind. The ecosystem is designed to avoid
divisions between classes of users with different interests. Crucially, there is no distinction between miners,35
developers, and users.
Governable and Democratic $AVAX is a highly inclusive platform, which enables anyone to connect to its
network and participate in validation and first-hand in governance. Any token holder can have a vote in
selecting key financial parameters and in choosing how the system evolves.
Interoperable and Flexible Avalanche is designed to be a universal and flexible infrastructure for a multitude40
of blockchains/assets, where the base $AVAX is used for security and as a unit of account for exchange. The
system is intended to support, in a value-neutral fashion, many blockchains to be built on top. The platform
is designed from the ground up to make it easy to port existing blockchains onto it, to import balances, to
support multiple scripting languages and virtual machines, and to meaningfully support multiple deployment
scenarios.45
Outline The rest of this paper is broken down into four major sections. Section 2 outlines the details of the
engine that powers the platform. Section 3 discusses the architectural model behind the platform, including
subnetworks, virtual machines, bootstrapping, membership, and staking. Section 4 explains the governance
model that enables dynamic changes to key economic parameters. Finally, in Section 5 explores various
peripheral topics of interest, including potential optimizations, post-quantum cryptography, and realistic50
adversaries.
Avalanche Platform 2020/06/30 3
Naming Convention The name of the platform isAvalanche, and is typically referred to as “theAvalanche
platform”, and is interchangeable/synonymous with “the Avalanche network”, or – simply – Avalanche.
Codebases will be released using three numeric identifiers, labeled “ v.[0-9].[0-9].[0-100]”, where the
first number identifies major releases, the second number identifies minor releases, and the third number55
identifies patches. The first public release, codenamed Avalanche Borealis , is v. 1.0.0 . The native token
of the platform is called “ $AVAX”. The family of consensus protocols used by the Avalanche platform is
referred to as the Snow* family. There are three concrete instantiations, called Avalanche, Snowman, and
Frosty.
2 The Engine60
Discussion of the Avalanche platform begins with the core component which powers the platform: the
consensus engine.
Background Distributed payments and – more generally – computation, require agreement between a set
of machines. Therefore, consensus protocols, which enable a group of nodes to achieve agreement, lie at the
heart of blockchains, as well as almost every deployed large-scale industrial distributed system. The topic65
has received extensive scrutiny for almost five decades, and that effort, to date, has yielded just two families
of protocols: classical consensus protocols, which rely on all-to-all communication, and Nakamoto consensus,
which relies on proof-of-work mining coupled with the longest-chain-rule. While classical consensus protocols
can have low latency and high throughput, they do not scale to large numbers of participants, nor are they
robust in the presence of membership changes, which has relegated them mostly to permissioned, mostly70
static deployments. Nakamoto consensus protocols [5, 7, 4], on the other hand, are robust, but suffer from
high confirmation latencies, low throughput, and require constant energy expenditure for their security.
The Snow* family of protocols, introduced by Avalanche, combine the best properties of classical con-
sensus protocols with the best of Nakamoto consensus. Based on a lightweight network sampling mechanism,
they achieve low latency and high throughput without needing to agree on the precise membership of the75
system. They scale well from thousands to millions of participants with direct participation in the consen-
sus protocol. Further, the protocols do not make use of PoW mining, and therefore avoid its exorbitant
energy expenditure and subsequent leak of value in the ecosystem, yielding lightweight, green, and quiescent
protocols.
Mechanism and Properties The Snow* protocols operate by repeated sampling of the network. Each node80
polls a small, constant-sized, randomly chosen set of neighbors, and switches its proposal if a supermajority
supports a different value. Samples are repeated until convergence is reached, which happens rapidly in
normal operations.
We elucidate the mechanism of operation via a concrete example. First, a transaction is created by
a user and sent to a validating node, which is a node participating in the consensus procedure. It is then85
propagated out to other nodes in the network via gossiping. What happens if that user also issues a conflicting
4 Kevin Sekniqi, Daniel Laine, Stephen Buttolph, and Emin G¨ un Sirer
transaction, that is, a doublespend? To choose amongst the conflicting transactions and prevent the double-
spend, every node randomly selects a small subset of nodes and queries which of the conflicting transactions
the queried nodes think is the valid one. If the querying node receives a supermajority response in favor
of one transaction, then the node changes its own response to that transaction. Every node in the network90
repeats this procedure until the entire network comes to consensus on one of the conflicting transactions.
Surprisingly, while the core mechanism of operation is quite simple, these protocols lead to highly
desirable system dynamics that make them suitable for large-scale deployment.
– Permissionless, Open to Churn, and Robust . The latest slew of blockchain projects employ classical
consensus protocols and therefore require full membership knowledge. Knowing the entire set of par-95
ticipants is sufficiently simple in closed, permissioned systems, but becomes increasingly hard in open,
decentralized networks. This limitation imposes high security risks to existing incumbents employing
such protocols. In contrast, Snow* protocols maintain high safety guarantees even when there are well-
quantified discrepancies between the network views of any two nodes. Validators of Snow* protocols
enjoy the ability to validate without continuous full membership knowledge. They are, therefore, robust100
and highly suitable for public blockchains.
– Scalable and Decentralized A core feature of the Snow family is its ability to scale without incurring
fundamental tradeoffs. Snow protocols can scale to tens of thousands or millions of nodes, without dele-
gation to subsets of validators. These protocols enjoy the best-in-class system decentralization, allowing
every node to fully validate. First-hand continuous participation has deep implications for the security105
of the system. In almost every proof-of-stake protocol that attempts to scale to a large participant set,
the typical mode of operation is to enable scaling by delegating validation to a subcommittee. Natu-
rally, this implies that the security of the system is now precisely as high as the corruption cost of the
subcommittee. Subcommittees are furthermore subject to cartel formation.
In Snow-type protocols, such delegation is not necessary, allowing every node operator to have a first-110
hand say in the system, at all times. Another design, typically referred to as state sharding, attempts
to provide scalability by parallelizing transaction serialization to independent networks of validators.
Unfortunately, the security of the system in such a design becomes only as high as the easiest corruptible
independent shard. Therefore, neither subcommittee election nor sharding are suitable scaling strategies
for crypto platforms.115
– Adaptive. Unlike other voting-based systems, Snow* protocols achieve higher performance when the
adversary is small, and yet highly resilient under large attacks.
– Asynchronously Safe . Snow* protocols, unlike longest-chain protocols, do not require synchronicity to
operate safely, and therefore prevent double-spends even in the face of network partitions. In Bitcoin,
for example, if synchronicity assumption is violated, it is possible to operate to independent forks of the120
Bitcoin network for prolonged periods of time, which would invalidate any transactions once the forks
heal.
– Low Latency. Most blockchains today are unable to support business applications, such as trading or daily
retail payments. It is simply unworkable to wait minutes, or even hours, for confirmation of transactions.
Therefore, one of the most important, and yet highly overlooked, properties of consensus protocols is the125
time to finality. Snow* protocols reach finality typically in ≤ 1 second, which is significantly lower than
both longest-chain protocols and sharded blockchains, both of which typically span finality to a matter
of minutes.
Avalanche Platform 2020/06/30 5
– High Throughput. Snow* protocols, which can build a linear chain or a DAG, reach thousands of transac-
tions per second (5000+ tps), while retaining full decentralization. New blockchain solutions that claim130
high TPS typically trade off decentralization and security and opt for more centralized and insecure
consensus mechanisms. Some projects report numbers from highly controlled settings, thus misreporting
true performance results. The reported numbers for $AVAX are taken directly from a real, fully imple-
mented Avalanche network running on 2000 nodes on AWS, geo-distributed across the globe on low-end
machines. Higher performance results (10,000+) can be achieved through assuming higher bandwidth135
provisioning for each node and dedicated hardware for signature verification. Finally, we note that the
aforementioned metrics are at the base-layer. Layer-2 scaling solutions immediately augment these results
considerably.
Comparative Charts of Consensus Table 1 describes the differences between the three known families
of consensus protocols through a set of 8 critical axes.140
Nakamoto Classical Snow*
Robust (Suitable for Open Settings) + - +
Highly Decentralized (Allows Many Validators) + - +
Low Latency and Quick Finality (Fast Transaction Confirmation) - + +
High Throughput (Allows Many Clients) - + +
Lightweight (Low System Requirements) - + +
Quiescent (Not Active When No Decisions Performed) - + +
Safety Parameterizable (Beyond 51% Adversarial Presence) - - +
Highly Scalable - - +
T able 1. Comparative chart between the three known families of consensus protocols. Avalanche, Snowman, and
Frosty all belong to the Snow* family.
3 Platform Overview
In this section, we provide an architectural overview of the platform and discuss various implementation
details. The Avalanche platform cleanly separates three concerns: chains (and assets built on top), execution
environments, and deployment.
3.1 Architecture145
Subnetworks A subnetwork, or subnet, is a dynamic set of validators working together to achieve consensus
on the state of a set of blockchains. Each blockchain is validated by one subnet, and a subnet can validate
arbitrarily many blockchains. A validator may be a member of arbitrarily many subnets. A subnet decides
who may enter it, and may require that its constituent validators have certain properties. The Avalanche
platform supports the creation and operation of arbitrarily many subnets. In order to create a new subnet150
or to join a subnet, one must pay a fee denominated in $AVAX.
6 Kevin Sekniqi, Daniel Laine, Stephen Buttolph, and Emin G¨ un Sirer
The subnet model offers a number of advantages:
– If a validator doesn’t care about the blockchains in a given subnet, it will simply not join that subnet.
This reduces network traffic, as well as the computational resources required of validators. This is in
contrast to other blockchain projects, in which every validator must validate every transaction, even155
those they don’t care about.
– Since subnets decide who may enter them, one can create private subnets. That is, each blockchain in
the subnet is validated only by a set of trusted validators.
– One can create a subnet where each validator has certain properties. For example, one could create a
subnet where each validator is located in a certain jurisdiction, or where each validator is bound by some160
real-world contract. This may be benificial for compliance reasons.
There is one special subnet called the Default Subnet. It is validated by all validators. (That is, in order
to validate any subnet, one must also validate the Default Subnet.) The Default Subnet validates a set of
pre-defined blockchains, including the blockchain where $AVAX lives and is traded.
Virtual Machines Each blockchain is an instance of a Virtual Machine (VM.) A VM is a blueprint for a165
blockchain, much like a class is a blueprint for an object in an object-oriented programming language. The
interface, state and behavior of a blockchain is defined by the VM that the blockchain runs. The following
properties of a blockchain, and other, are defined by a VM:
– The contents of a block
– The state transition that occurs when a block is accepted170
– The APIs exposed by the blockchain and their endpoints
– The data that is persisted to disk
We say that a blockchain “uses” or “runs” a given VM. When creating a blockchain, one specifies the VM
it runs, as well as the genesis state of the blockchain. A new blockchain can be created using a pre-existing
VM, or a developer can code a new one. There can be arbitrarily many blockchains that run the same VM.175
Each blockchain, even those running the same VM, is logically independent from others and maintains its
own state.
3.2 Bootstrapping
The first step in participating in Avalanche is bootstrapping. The process occurs in three stages: connection
to seed anchors, network and state discovery, and becoming a validator.180
Seed Anchors Any networked system of peers that operates without a permissioned (i.e. hard-coded)
set of identities requires some mechanism for peer discovery. In peer-to-peer file sharing networks, a set of
trackers are used. In crypto networks, a typical mechanism is the use of DNS seed nodes (which we refer
Avalanche Platform 2020/06/30 7
to as seed anchors), which comprise a set of well-defined seed-IP addresses from which other members of
the network can be discovered. The role of DNS seed nodes is to provide useful information about the set185
of active participants in the system. The same mechanism is employed in Bitcoin Core [1], wherein the
src/chainparams.cpp file of the source code holds a list of hard-coded seed nodes. The difference between
BTC and Avalanche is that BTC requires just one correct DNS seed node, whileAvalanche requires a simple
majority of the anchors to be correct. As an example, a new user may choose to bootstrap the network view
through a set of well established and reputable exchanges, any one of which individually are not trusted.190
We note, however, that the set of bootstrap nodes does not need to be hard-coded or static, and can be
provided by the user, though for ease of use, clients may provide a default setting that includes economically
important actors, such as exchanges, with which clients wish to share a world view. There is no barrier to
become a seed anchor, therefore a set of seed anchors can not dictate whether a node may or may not enter
the network, since nodes can discover the latest network of Avalanche peers by attaching to any set of seed195
anchors.
Network and State Discovery Once connected to the seed anchors, a node queries for the latest set of
state transitions. We call this set of state transitions the accepted frontier. For a chain, the accepted frontier
is the last accepted block. For a DAG, the accepted frontier is the set of vertices that are accepted, yet have
no accepted children. After collecting the accepted frontiers from the seed anchors, the state transitions that200
are accepted by a majority of the seed anchors is defined to be accepted. The correct state is then extracted
by synchronizing with the sampled nodes. As long as there is a majority of correct nod프로젝트로 돌아가기