bitpro가 원문에서 글자만 모은 내용이에요. 표·그림·수식은 빠지거나 깨질 수 있어요. 전체 2쪽 중 2쪽
s defined by the client) as needed based on
any number of input metrics. The "management tier" may be spr ead
across all datacenters for a deployment, with global state m aintained
by a distributed database running over the secure overlay ne twork.
5.1 Example: Latency-Optimized Deployment
Figure 4: Illustration of slower performance due to higher l atencies for end-
users distributed across the globe for a single datacenter d eployment
Many web-based applications are latency-sensitive - lower response
times from application servers translates into a dramatica lly improved
end-user experience. Modern deployments of such applicati ons employ
content delivery networks (CDNs) to deliver static content such as
images to end users quickly.
CDNs provide reduced latency by distributing content so tha t it is
geographically close to the users that are accessing it. Dep loyments on
the Akash Network can not only replicate this approach, but b eat it
- Akash gives clients the ability to place dynamic content cl ose to an
application’s users.
To implement a self-managed dynamic delivery network on Akash, a
DevOps engineer would include a management tier in their dep loyment
which monitors the geographical location of clients. This m anagement
12
Figure 5: Illustration of improved network performance by d ynamically dis-
tributing workloads and their state across datacenters in c lose proximity to
the end-users
tier would add and remove datacenters across the globe, prov isioning
more resources in regions where user activity is high, and le ss resources
in regions where user participation is low.
5.2 Example: Machine Learning Deployment
Machine learning applications employ a large number of node s to par-
allelize computations involving large datasets. They do th eir work in
"batches" - there is no "steady state" of capacity that is req uired.
A machine learning application on Akash may use a management
tier to proactively procure resources within a single datac enter. As a
machine learning task begins, the management tier can "scal e up" the
number of nodes for it; when a task completes, the resources p rovi-
sioned for it can be relinquished.
13
MASTER WORKER
DATACENTER
Figure 6: A machine learning batch job under less load runnin g a single
master and single worker node
MASTER
WORKER
WORKER
WORKER
DATACENTER
Figure 7: A machine learning batch job under load running a si ngle master
and multiple worker nodes
14
References
[Amazon(2017a)] Amazon. Amazon elasticache pricing. 2017 a. URL
https://aws.amazon.com/elasticache/pricing/.
[Amazon(2017b)] Amazon. Amazon ec2 pricing. 2017b. URL
https://aws.amazon.com/ec2/pricing/.
[Cisco(2016)] Cisco. Cisco global cloud index: Fore-
cast and methodology, 2015 - 2020. 2016. URL
https://www.cisco.com/c/dam/en/us/solutions/collateral/service-provider/global-cloud-
[Forrester(2017)] Forrester. Predictions 2018: Cloud com puting
accelerates enterprise transformation everywhere. 2017. URL
https://www.forrester.com/report/Predictions+2018+Cloud+Computing+Accelerates+Enterpri
[Gartner(2017)] Gartner. Forecast analysis: Public cloud
services, worldwide, 2q17 update. 2017. URL
https://www.gartner.com/doc/3803517.
[Glanz(2012)] James Glanz. Power, pollu-
tion and the internet. 2012. URL
http://www.nytimes.com/2012/09/23/technology/data-c enters-waste-vast-amounts-of-energy
[Kaplan et al.(2008)Kaplan, Forrest, and Kindler] James Ka -
plan, William Forrest, and Noah Kindler. Revolu-
tionizing data center energy efficiency. 2008. URL
https://www.sallan.org/pdf-docs/McKinsey_Data_Cente r_Efficiency.pdf.
[Koomey and Taylor(2015)] Jonathan Koomey and Jon Tay-
lor. New data supports finding that 30 percent of servers
are ’comatose’, indicating that nearly a third of capi-
tal in enterprise data centers is wasted. 2015. URL
https://anthesisgroup.com/wp-content/uploads/2015/0 6/Case-Study_DataSupports30PercentC
[Liu(2011)] Huan Liu. A measurement study of
server utilization in public clouds. 2011. URL
http://ieeexplore.ieee.org/document/6118751/.
15프로젝트로 돌아가기