bitpro가 원문에서 글자만 모은 내용이에요. 표·그림·수식은 빠지거나 깨질 수 있어요. 전체 2쪽 중 1쪽
Akash Network: Decentralized Cloud
Infrastructure Marketplace
Overclock Labs
May 13, 2018
Version 0.0.3
Note: The Akash Network is an active research project and new v ersions of this
paper will manifest at akash.network. For comments and sugge stions please reach
out at research@akash.network
Abstract
Cloud computing — the process of offloading work to remote serv ers
– is inherently broken. While it mostly works as advertised, we’ve
found that inefficiencies still plague the system. The produc ts pro-
duced by the major cloud providers are usable but they are lim ited
to shortcomings that can be solved today with advancements i n con-
tainer technology and a powerful token economy. The purpose of this
white paper is to put forward our plan for a cloud services mar ket
called Akash Network, the worlds first global spot market for cloud
computing.
We see a future where the global cloud infrastructure of the w orld
is decentralized and distributed between all cloud service providers; a
market that deploys and liquidates (increasingly commodit ized) data
center compute in a secure, fast and transparently spot pric ed manner.
Services are sold in a democratic but unified ecosystem that a nyone
can use.
In this paper, we present Akash, cloud infrastructure netwo rk that
is decentralized, competitive, and able to distribute appl ications be-
tween multiple cloud service providers around the globe. Th e paper
will introduce the state of the existing market, outline how we are
using latest developments in serverless container orchest ration to com-
bat these issues, the basics of and the necessity of the netwo rks native
token, AKASH, and finally our roadmap for launch.
1
Contents
1 Introduction 4
1.1 A Troubled Industry . . . . . . . . . . . . . . . . . . . . 4
2 The Akash Network 6
2.1 The Akash Blockchain . . . . . . . . . . . . . . . . . . . 6
2.2 The Akash Token, AKASH . . . . . . . . . . . . . . . . 7
3 Marketplace 8
4 Deployment 10
4.1 Manifest Distribution . . . . . . . . . . . . . . . . . . . 10
4.2 Overlay Network . . . . . . . . . . . . . . . . . . . . . . 10
5 Automation 11
5.1 Example: Latency-Optimized Deployment . . . . . . . . 12
5.2 Example: Machine Learning Deployment . . . . . . . . . 13
2
List of Figures
1 Illustration of on-chain and off-chain interactions among st
various participants in the Akash network . . . . . . . . 7
2 Summary of procurement from Marketplace. (1) User’s
deployment order is posted to the orderbook (2) Dat-
acenters posts eligible fulfillment orders for the deploy-
ment order (3) The best fulfillment order is matched
with the deployment order, creating a new lease. . . . . 9
3 Illustration of Akash’s overlay network . . . . . . . . . . 11
4 Illustration of slower performance due to higher laten-
cies for end-users distributed across the globe for a single
datacenter deployment . . . . . . . . . . . . . . . . . . . 12
5 Illustration of improved network performance by dy-
namically distributing workloads and their state across
datacenters in close proximity to the end-users . . . . . 13
6 A machine learning batch job under less load running a
single master and single worker node . . . . . . . . . . . 14
7 A machine learning batch job under load running a sin-
gle master and multiple worker nodes . . . . . . . . . . 14
3
1 Introduction
The Akash Network (Akash) is a secure, transparent, and decentralized
cloud computing marketplace that connects those who need co mputing
resources ( clients) with those that have computing capacity to lease
(providers).
Akash acts as a "super" cloud platform ( supercloud) - providing a
unified layer above all providers on the marketplace so as to p resent
clients with a single cloud platform, regardless of which pa rticular
provider they may be using.
Clients use Akash because of its cost advantage, usability, and flex-
ibility to move between cloud providers, and the performanc e benefits
of global deployments. Providers use Akash because it allow s them to
earn profits from either dedicated or temporarily-unused ca pacity.
1.1 A Troubled Industry
By 2020, cloud infrastructure providers will account for 53% of global
internet traffic[Cisco(2016)], out of which Amazon, Google, and Mi-
crosoft will deliver 80% of the payload[Forrester(2017)].
While the cloud will deliver the majority of the workloads, t he
future of the internet stands at a risk of being consolidated , centralized,
and at the mercy of these three providers.
The primary driver for cloud adoption is the promise of flexib ility
and cost advantage, but the reality is that the products offer ed by cloud
providers are overpriced, complicated, and lock clients in to ecosystems
that limit their ability to innovate, compete, and have sove reignty over
their infrastructure needs.
The difference in capital expenditure of purchasing hardwar e and
leasing datacenters between running in the cloud and self ma naging
(on-premise) is marginal; however, the cloud providers hav e a signif-
icant advantage with operating expenditure because of thei r invest-
ments in automation with minimal human touch.
Even though running computing on-premise can offer much bet-
ter flexibility, performance, and security, organizations are abandoning
their datacenter operations and migrating to the cloud beca use they
are finding it increasingly hard to justify the operating cos ts due to lack
of adequate automation along with low utilization footprin t. Idle, un-
derutilized servers prove to be costly and wasteful. Analys ts estimate
that as many as 85% of servers in practice have underutilized capacity
[Glanz(2012)] [Kaplan et al.(2008)Kaplan, Forrest, and Ki ndler] [Liu(2011)]
[Koomey and Taylor(2015)].
Cloud providers drive margins by building hyper-scale inst allations,
i.e, consolidating resources in few datacenters for econom ic efficiency,
4
and cross-selling fully managed backend services, such as d atabases,
cache stores, API gateways, etc.
Being hyper-scale allows them to oversubscribe their custo mers,
hence driving higher margins but creates single-points for failures. Ge-
ographically distributed workloads offer much reliability and end-user
performance; however, the cloud providers make it extremel y hard for
clients to be multi-regional because it doesn’t work in thei r best inter-
est.
The cloud providers prefer customers to deploy their applic ations
in a single datacenter and penalize them for being cross-reg ional or
multi-zonal, usually through hefty bandwidth fees and vari able regional
pricing. This is why A WS’ pricing model is different for each r egion
for the same exact resource.
Even though selling instances is lucrative, Cloud Provider s usually
charge a small amount for instances compared to the premium t hey
charge for managed backend services ( PaaS ); analogous to the old
burgers-and-fries model where a restaurant needs to sell bu rgers at a
loss so that they can sell the more addictive fries at a high ma rgin.
The PaaS services sold by the providers tend to be white-labe led
open source projects where the original authors are never in centivized,
and the cloud providers have no incentive to evolve the produ ct. For
example, A WS’ ElastiCache is a white-labeled open source so ftware
called Redis. Redis is an open source project — much loved by d e-
velopers — written by Salvatore Sanfilippo and maintained by Redis
Labs.
As of the writing of this paper, a managed Redis server, in US E ast
(Ohio) running on r3.8xlarge is priced at $31,449/yr [Amazon(2017a)]
whereas the same instance without Redis costs $18,385/yr [A mazon(2017b)].
The extra $13,064 just for a "piece of mind" to the customer. N either
Sanfilippo or Redis Labs are incentivized for the efforts.
Also, more services mean more dependent the customer is on th e
cloud provider. The complexity introduced by increasing am ounts of
features, service availability, and codification using non -standard APIs
lead to customers being locked in by the cloud vendors, preve nting
clients from exploring other better options in the marketpl ace while
inhibiting innovation.
This model adopted by the providers stifles innovation as it d ramat-
ically reduces the chance of an open source project from succ eeding.
Cloud providers effectively act as middle-men that set the ru les of en-
gagement for the industry while making a no contribution to s ociety
on the whole.
5
2 The Akash Network
The foundational design objective of the Akash Network is to maintain
a low barrier to entry for providers while at the same time ens uring that
clients can trust the resources that the platform offers them . To achieve
this, the system requires a publicly-verifiable record of tr ansactions
within the network. To that end, the Akash Network is impleme nted
using blockchain technologies as a means of achieving conse nsus on the
veracity of a distributed database.
Akash is, first and foremost, a platform that allows clients t o pro-
cure resources from providers. This is enabled by a blockcha in-powered
distributed exchange where clients post their desired reso urces for
providers to bid on. The currency of this marketplace is a dig ital
token, the Akash (AKASH), whose ledger is stored on a blockch ain.
Akash is a cloud platform for real-world applications. The r equire-
ments of such applications include:
• Many workloads deployed across any number of datacenters.
• Connectivity restrictions which prevent unwanted access t o work-
loads.
• Self-managed so that operators do not need to constantly ten d
to deployments.
To support running workloads on procured resources, Akash i n-
cludes a peer-to-peer protocol for distributing workloads and deploy-
ment configuration to and between a client’s providers.
Workloads in Akash are defined as Docker containers. Docker c on-
tainers allow for highly-isolated and configurable executi on environ-
ments, and are already part of many cloud-based deployments today.
2.1 The Akash Blockchain
The Akash blockchain provides a layer of trust in a decentral ized and
trustless environment. Clients inherently trust today’s l arge infras-
tructure Providers based primarily on the brand equity they ’ve built
over years.Akash does not and should not require that same le ap of
faith, since any Provider with capacity can compete to offer s ervices
on Akash. Instead, the blockchain earns trust via an open and trans-
parent platform. Data on the chain is an immutable and public record
of all transactions, including each Provider’s fulfillment history.
Akash is also politically decentralized. No single entity c ontrols
the network and no intermediary facilitates transactions. Therefore no
entity is incentivized to control or to extract marginal rev enue from
the network. As an example, a large company such as Coca-Cola can
6
Photon Blockchain
Delegate Delegate
Delegate
consensus
new blocks
Client
Provider
Provider
Provider
peer-to-peer
private overlay network
End
User
http / tcp
Figure 1: Illustration of on-chain and off-chain interactio ns amongst various
participants in the Akash network
participate in the network as a Provider, providing compute to another
large company or to an individual developer, yet all three pa rties are
on equal footing in the network.
2.2 The Akash Token, AKASH
The Akash Token (AKASH) is used to simplify the exchange of va lue
and align economic incentives with proper user behavior. Th e Akash
token is the marketplace currency used to pay for leased comp ute in-
frastructure on Akash’s decentralized network. Our token s erves two
primary functions in Akash’s ecosystem.
In a market that is expected to be $737 billion , with well over
21% annual growth [Gartner(2017)], the liquidity of AKASH w ill be
matched by the demand for compute power. Along this line of th ought,
7
we have full confidence in the network and for AKASH to achieve
maximum liquidity for its early adopters and end state user.
2.2.1 Staking
The stability of the Akash network relies on a staking system that pre-
vents bad actors from abusing our system. A staking system pr ovides
a prohibitive monetary disincentive for bad actors who cons ider par-
ticipating in our network. The risk of fraudulent behavior i s highest
when new, unknown providers join our network. Rather than re quir-
ing a centralized or federated approval process for new acco unts, the
Akash network allows anyone to join.
When a new provider chooses to offer its resources on the Akash
network, rather than being approved, it must stake a meaning ful value
on the network in Akash tokens. There is no minimum stake amou nt,
but participation in Akash Network governance is proportio nal to a
providerâĂŹs stake, taken as a fraction of the sum of all stak es. Ad-
ditionally, stake contribution is factored into a provider ’s reputation
score, which tenants may use as a deployment criterion.
2.2.2 Global Payments
Akash tokens mitigate the foreign exchange risk that usuall y results
from cross-border payments. Taking the place of fiat for thes e transac-
tions, Akash tokens simplify the exchange of value in the clo ud infras-
tructure industry. Our matching engine competitively pric es each con-
tainer compute against a prevailing market amount of Akash t okens.
When a tenant is matched with a provider, the tenant pays Akas h
tokens to the network, which are subsequently paid to the pro vider
according to the terms of the lease.
3 Marketplace
Infrastructure procurement — the process through which cli ents lease
infrastructure from providers — on Akash is implemented thr ough a
decentralized exchange ( marketplace).
The marketplace consists of a public order book and a matchin g al-
gorithm. Clients place deployment orders, which contain a specification
of the client’s service needs, and datacenters place fulfillment orders to
bid on deployment orders. Deployment orders include the max imum
amount the client is willing to pay for a fixed number of comput ing
units (as measured by memory, cpu, storage, and bandwidth) f or a
specific amount of time; fulfillment orders declare the price that the
provider will provide the resources for.
8
Deployment orders are open for a client-defined length of time, as
measured to the second. While the deployment order is open, p roviders
may post fulfillment orders to bid on it.
A fulfilment order is eligible to match with a deployment orde r if
the fulfillment order satisfies all minimum specifications of the deploy-
ment order. Given a deployment order and a set of eligible ful filment
orders, the fulfilment order offering the lowest price will be matched
with the deployment order. If multiple fulfilment orders are eligible for
a match and offer the same price, the fulfilment order placed fir st will
be matched with the deployment order.
Businesses and individual consumers will want and need to pr otect
how they are publicly displaying their use of compute power. To guard
against competitor data mining and other attack vectors, a h omomor-
phic encryption layer is added.
A lease is created when a match occurs between a deployment and
fulfillment order. The lease contains references to the depl oyment and
fulfilment orders. Leases will be the binding agent in fulfill ing a de-
ployment.
user orderbook 1
lease
3
datacenter
2
datacenter 2
datacenter
2
Figure 2: Summary of procurement from Marketplace. (1) User ’s deploy-
ment order is posted to the orderbook (2) Datacenters posts e ligible ful-
fillment orders for the deployment order (3) The best fulfillm ent order is
matched with the deployment order, creating a new lease.
9
4 Deployment
Once resources have been procured, clients must distribute their work-
loads to providers so that they can execute on the leased reso urces.
We refer to the current state of the client’s workloads on the Akash
Network as a deployment.
A user describes their desired deployment in a manifest. The man-
ifest is written in a declarative file format that contains wo rkload def-
initions, configuration, and connection rules. Providers u se workload
definitions and configuration to execute the workloads on the resources
they are providing, and use the connection rules to build an o verlay
network and firewall configurations.
A hash of the manifest is known as the deployment version and is
stored on the blockchain-based distributed database.
4.1 Manifest Distribution
The manifest contains sensitive information which should only be shared
with participants of the deployment. This poses a problem fo r self-
managed deployments - Akash must distribute the workload de finition
autonomously, without revealing its contents to unnecessa ry partici-
pants.
To address these issues, we devised a peer-to-peer file shari ng scheme
in which lease participants distribute the manifest to one a nother as
needed. The protocol runs off-chain over a TLS connection; ea ch par-
ticipant can verify the manifest they received by computing its hash
and comparing this with the deployment version that is store d on the
blockchain-backed distributed database.
In addition to providing private, secure, autonomous manif est dis-
tribution, the peer-to-peer protocol also enables fast dis tribution of
large manifests to a large number of datacenters.
4.2 Overlay Network
By default, a workload’s network is isolated - nothing can co nnect to
it. While this is secure, it is not practical for real-world a pplications.
For example, consider a simple web application: end-user br owsers
should have access to the web tier workload, and the web tier n eeds
to communicate to the database workload. Furthermore, the w eb tier
may not be hosted in the same datacenter as the database.
On the Akash Network, clients can selectively allow communi ca-
tions to and between workloads by defining a connection topol ogy
within the manifest. Datacenters use this topology to config ure firewall
rules and to create a secure network between individual work loads as
needed.
10
APP
APP
health checker
APP
health checker health checker
shared state replicated
using a private
peer-to-peer overlay
network
Datacenter Datacenter
Datacenter
Figure 3: Illustration of Akash’s overlay network
To support secure cross-datacenter communications, provi ders ex-
pose workloads to each other through a mTLS tunnel. Each work load-
to-workload connection uses a distinct tunnel.
Before establishing these tunnels, providers generate a TL S certifi-
cate for each required tunnel and exchange these certificate s with the
necessary peer providers. Each provider’s root certificate is stored on
the blockchain-based distributed database, enabling peer s to verify the
authenticity of the certificates it receives.
Once certificates are exchanged, providers establish an aut henti-
cated tunnel and connect the workload’s network to it. All of this
is transparent to the workloads themselves - they can connec t to one
another through stable addresses and standard protocols.
5 Automation
The dynamic nature of cloud infrastructure is both a blessin g and a
curse for operations management. That new resources can be p rovi-
sioned at will is a blessing; the exploding management overh ead and
complexity of said resources is a curse. The goal of DevOps — the
practice of managing deployments programmatically — is to a lleviate
the pain points of cloud infrastructure by leveraging its st rengths.
The Akash Network was built from the ground up to provide De-
vOps engineers with a simple but powerful toolset for creati ng highly-
automated deployments. The toolset is comprised of the prim itives
that enable non-management applications — generic workloa ds and
11
overlay networks — and can be leveraged to create autonomous , self-
managed systems.
Self-managed deployments on Akash are a simple matter of cre ating
workloads that manage their own deployment themselves. A De vOps
engineer may employ a workload that updates DNS entries as pr oviders
join or leave the deployment; tests response times of web tie r appli-
cations; and scales up and down infrastructure (in accordan ce with
permissions and constraint프로젝트로 돌아가기