Distributed Computing with Idle Resources · 2019-03-28 · On-going trends: • IoT and fog...
Transcript of Distributed Computing with Idle Resources · 2019-03-28 · On-going trends: • IoT and fog...
AnkrDistributed Computing with Idle Resources
Ankr strives to leverage idle
processing power in data centers and
connected devices for a distributed,
secure and intuitive cloud.
“ “
I. Cloud evolution and the problem today
II. Distributed cloud enablers - why now
III. Ankr technology and milestones
IV. “Sky is the limit” - rapid go-to-market strategy
V. Team behind the distributed Cloud
I. Cloud evolution and the problem today
On-going trends:
• IoT and fog computing (geographically distributed data sources)
Pharma/bio/chemistry Industrial research
Engineering & manufacturing simulation
Film/AR/VR rendering
• Growth of serverless computing (focus on computational task, not computational machine)
Serverless architecture
Distributed data (Bringing compute and data closer to the users)
Rising market needs for large processing power:
History of computing
Mainframes
On-premises data center
Private cloud (e.g., Rackspace)/colocation data centers
Public cloud(e.g., AWS, Azure, Alibaba)
Whats next?
Cloud evolution and customer trends
Low utilization rate of data centersAccording to research from McKinsey and Stanford,
• Roughly 30% of servers (including standard server, hypervisors and VMs) in centers around the world have showed no signs of network, user, connection, memory or CPU activity in six months or more
• Roughly 35% of servers have showed occasional signs of activity (compute and network) <5% of the time
• In total, near 70% “non-performing” server asset!
Problems: Data centers aren’t performing Opportunities: Win-Win
Data centers aren’t able to compete
with the public cloud
• Users trust large CSPs (AWS, Azure, Digital Ocean, AlibabaCloud) due to their wide brand
awareness & recognition, while they trust small data centers without brand names much less
• AWS & other large CSPs exploit economies of scale and network effects
• AWS & other public cloud provide similar, universal interfaces that engineers find easy to use
• Pay-as-you-go model (eg. Lambda) charges higher unit-time price, but lower total cost
Idle data centers are long-term headaches
• Companies can’t entirely sell their data centers, using them internally for specific needs
• Data centers require large CapEx for companies, & underperforming servers are
deadweight with annual depreciation lost, adding pressure on companies’ bottom line
• Public cloud provider only offers to buy up the entire premise for low price
• Rental agreements for partial access to data centers are big-ticket slow sales, and
integration for renter causes tremendous overhead concerns
Elastically utilize local data centers for cloud computation is a win-win
• Prevailing low utilization rate of data centers means large computation potential
• Data center owners can keep certain application/data on-premise
• Elastic computation resource utilization will outperform traditional data center broker model, solving resource planning and integration problems
• Ankr’s containerized solution can be deployed to data centers in just hours
• Additional flexible revenue stream for data center owners
• Local distributed cloud can deliver favorable performance given the growth of IoT and local data sources
For consumers:Ankr uses local data centers to offer fast & cheap computation
For small data centers:Ankr offers new revenue stream for under-utilized data centers
Problems and opportunities
II. Distributed Cloud – Why Now?
Enabler: containerization & cluster orchestration
Docker was released and quickly became
the most popular container
Originated from Google Borg, K8’s container
orchestration has been widely adopted
Apache Mesos takes isolation to the next
step, supporting both VMs and containers
Kubernetes-as-a-Service has seen rapid growth on mainstream
cloud providers
Linux Container (LXC) was introduced for
running multiple Linux systems
VMs are known for better
security control.
2009 2013 2014 2016 2018
Large enterprises are abandoning legacy code in their monolithic applications as cost to manage and upgrade piles up
Microservices leveraging IaaS and PaaS can be deployed in different languages and scaled independently
Faas significantly reduces cost in auto-scaling compared to microservices
Virtual Machines vs. Containers Trend: Monolithic application →Microservices → FaaS
Virtual Machines
Containers Containers are lightweight, virtualizing only the
Operating System and take only seconds to start
Security Speed Management
VMs incur slower and more complex DevOps resource
management
VMs take MUCH more system resources (CPU and RAM), &
take minutes to start
However, containers enable MUCH less code migration and robust security updates
The underlying orchestration technology enable easier and
faster DevOps resource management
Enabler: TEE (trusted execution environments)
Intel SGX was built into 6th generation
processors
NVIDIA introduced open source stack for
TEE
AMD SEV was started for EYPC processor line
Intel SGX 2 will be released
ARM TrustZone’s hardware-based
isolation launched
2014 2015 2016 2016 2018
Not achieved yet after decades of research work & very slow, even when it’ll be achieved Will approach untrusted computing performance
Fully homomorphic encryption (FHE) TEEs are engineering solution equivalent to FHE
Encrypted data in, encrypted data out
Protects confidentiality and integrity of data during computation
Enclave application guarantees confidentiality and integrity
Encrypted data in, decrypted data out
III. Ankr technology and milestones
Ankr distributed cloud for idle data centers
Orchestration and containerization make Ankr distributed cloud computing network easy to
integrate with resources from diverse, independent data centers into one homogeneous cloud
Enabling technologies
Docker is the de-facto standard technology for containerization of applications into micro-
services to be deployed on Kubernetes clusters
Kubernetes is a general-purpose vendor-agnostic orchestration service that makes it easy to deploy
containerized applications in any data center
Cost reduction enabled by distributed cloud
Cost reduction for Distributed Cloud
Single Server*
42 Servers
*Dell R320 as sample
$55 Server
$29 Power
$80 Space
$1
64
/mo
nth
$2
00
/mo
nth
$8
,40
0/m
on
th
$55*42 Servers
$29*42 Power
$80 Space + $27 Rack
$3
,63
5/m
on
th
($55*42)+$80 = ($2390) lost/month
when servers arenot utilized
Assumptions(per month):
• Nominal cost = $55• Power usage costs = $29
520.8kWh energy consumptionPower supplies pull 30% and are 80% efficient1.7 PUE average265.2 kWh actual power consumption$0.11 average commercial price per kWh in US
• Total location cost = $80$10 per square footServer size of 8 square feet
Distributed Cloud Cash Flow Advantages:
Lose money (up to $2,390) each month on
idle servers
AWS, using same set of servers, create a $8,400
revenue potential
Flexible structure enables flexible pricing
between $30-$200
• Challenges can be issued such that responding to to them correctly is guaranteed to necessitate actual work
• Solution to the delegated task can be quickly and verifiably reconstructed from the workers’ response
• Able to utilize otherwise wasteful work
Ankr distributed cloud for idle connected devices
PouW vs. PoW
• Work and energy expended to prove that work had in fact been done
• Mining requires highly specialized computer software to run complicated algorithms
• Computations guarantee the security of the network but cannot be applied to any other fields
• Providers and consumers run a daemon to interact with the network• Consumer uploads computational task (e.g., SGX enclave application)
to the Ankr network• Providers compete to be assigned the computational task• Select provider executes the task using its computational resources• Consumer verifies the results (e.g., SGX remote attestation) and pays
the provider
Peer-to-peer network of computational resource providers and consumers
Technical milestones & next steps
Milestones
Next Steps
Q4 2017 - Q1 2018
Research and theoretical framework validation
May 2018
Distributed cloud for connected devices with TEE proof of concept
July 2018
Distributed cloud for connected devices with TEE prototype
Q1 2019
Distributed cloud for idle data center resources proof of concept
September 2018
Distributed cloud for connected devices with TEE test network
October 2018
Distributed cloud for idle data center resources system design
Seek strategic corporate partner with data centers and pilot customers to achieve data center sharing end-to-end proof-of-concept and customer feedback
• Deploy applications using CLI to one of available data centers
• Ankr Hub manages the contributing data centers
Launch corporate-facing Initial service
• Elastic computation resource service (simulation, rendering and modeling)
• Accounting and payments for server providers and customers
Launch additional services
• Storage
• Data analytics / machine learning
• Content Delivery Network
IV. “Sky is the limit” - rapid go-to-market strategy
Analytics 4 Life’s product is an AI-powered medical imaging and
mathematical modelling platform.
• Tremendous spending on cloud
• Not enough cloud DevOps
• Hard-to-use existing platform: “I wish there’s a cloud button.”
HTC Vive produces VR Headsets. Dr. Liu is a data scientist and a Kaggle master.
• Existing cloud infrastructure hard to configure and expensive
• Often ended up just using university resources which due to its nature are bad consumer facing
Use Cases:
Monte Carlo simulations (e.g., Pharma/bio/chemistry Industrial
research)
Time-sensitive signal processing (offloading) (e.g., rendering for AR/VR)
Offline data analytics without deadlines
Content delivery from large data center to end-user (mobile) device
Internet-of-Things data collection from remote sensors to large data center
Use cases and future customer interviews
Analytics 4 Life HTC Vive Dr. Liu
● Significant overhead affecting user experiences
● Mostly compute-heavy applications
Asset-light sharing economy has proven its expanding speed
Airbnb Uber
Airbnb’s4 million listings
worldwide
More listings than the top 5 major hotel brand combined
Airbnb’s valuation compared to leading hotel brands (2017)
$9.3 B
$21 B
$31 B
$39 B
Number of Airbnb Guest Arrivals per Year
6 M16 M
40 M
80 M100 M
Projected Growth:
2016Revenue: $1.7 BProfit: $0.1 B
2017Revenue: $2.8 BProfit: $0.45 B
2020Revenue: $8.5 BProfit: $3.5 B
“60% of consumers who
have used Airbnb prefer it over
traditional hotels”
Uber is the first company to reach a $41 billion market capitalization in less than 6 years
Uber is valued higher than these 3 famous companies:
21st Century Fox$52.4 B
GM$49.9 B
Paypal$51.7 B
Uber’s Global Quarterly Gross Bookings:
$5.4 B$6.9 B $7.5 B $8.7 B $9.7 B $11.1 B $11.3 B
$12.0 B
75 M Riders
Uber’s by the numbers (2018):
3 MDrivers
10 B Total Rides
15 MDaily Rides
Uber’s 2018 Valuation: $72 Billion
Active Community Testing
Reputable Venture Funding
After the launching of Ankr’s MVP there are already
400 developer downloads
across the globe and we have cultivated an active developer
community for testing feedback
Shanghai Jiao Tong University (SJTU): Leading research institute in
Computer Science in the world
Work with SJTU’s departments to research and develop Ankr’s core components:
• Institute of Parallel & Distributed Computing
• Institute of Cryptography & Information SecurityDistinguished Alumni: Stanley Wu (CTO of Ankr)
(graduated in top 1% of class from SJTU)
Current focus:
• Test our functionalities using existing technology like SAP Hana
• Deploy functionalities on the SAP App Center
• Reach wide global enterprise audience
Ongoing projects:
• Using BOINC to reach a wider audience with scientific & academic backgrounds
• BOINC’s strong awareness is shown by its 200k+ contributors across the globe
• Caters to users with high likelihood of adopting Ankr
Partnership advantage:
Strong global awareness & active community testing
Berkeley Open Infrastructure for Network Computing
Advisor: Professor David Anderson
(creator of BOINC project and SETI@Home)
Leading Enterprise Software and Cloud Service Provider
V. Team behind the distributed Cloud
Prof. David AndersonTechnical Advisor
• Creator of BOINC and SETI@home• Researcher at Berkeley Space Science Lab
Research Engineering Product & Growth
Dr. Giacomo GhidiniChief Scientist
• NSF Grant Recipient • Assistant Professor at University of Texas
Arlington• Researcher at Oracle• Expert in blockchain and IoT
Dr. Wenli ZhengResearcher
• Assistant Professor at Shanghai Jiao Tong University
• Research Assistant at Ohio State University• Expert in computer architecture and data center
infrastructure
Dr. Quan ChenResearcher
• Assistant Professor at Shanghai Jiao Tong University
• Post-doc Researcher at Columbia University and University of Michigan Ann Arbor
• Expert in distributed systems
Stanley WuCTO
• 11 years at Amazon as Tech Lead and L6 Engineer
• Among the first few engineers to join AWS EC2• Expert in large-scale cloud service
Song LiuChief Security Engineer
• Principal Engineer at Palo Alto Networks and Gigamon
• Ethical Hacker and active open source contributor
• Presented multiple security vulnerabilities to Microsoft
Ambarish KrishnamurthyChief Architect
• VP Engineering at Morgan Stanley and Bank of America
• Led multiple ultra-low-latency trading systems • 25 years of experience in C/C++
Chandler SongCEO
• Serial Entrepreneur: Led the mobile product of CitySpade, an on-demand real estate broker service
• Software Engineer at Amazon Lab126 and SAP
Ryan FangCOO
• Serial Entrepreneur: Sold Peer Potential, an education competition start-up
• Investment Banking Analyst at Morgan Stanley and Credit Suisse
Appendix
Appendix: Current public cloud offerings
Centralized public cloud service providers (CSPs)
Cloud computing basic services
Compute
Storage: block and object
Networking: firewalls, caching, etc.
Access control, management, billing
Other services built on top of basic services
Big Data (e.g., AWS Elastic MapReduce)
Machine learning (e.g., MXNet on AWS)…
New computing paradigm
Serverless computing (e.g. AWS Lambda)
Current scenario Elastic compute pricing (AWS, Alibaba, Azure)
Serverless computing pricing
AWST3.medium (2 vCPUs, 4 GiB)
Alibabaecs.n4.large (2 CPUs, 4 GB)
AzureB2S (2 cores, 4 GiB)
All three providers feature similar pricing for elastic compute products
$0.0416/hr
$0.047/hr
$0.047/hr
Azure Functions
AWS Lambda
Alibaba Function Compute
All three providers feature same pricing for serverless products
$0.20/M executions$0.000016/GB-s duration
$0.20/M requests (i.e., executions)$0.00001667/GB-s duration
$0.20/1 million calls (i.e., executions)$0.00001668/GB-second duration fee
DreamLab’s Historic Success
Lacked Incentives
Vodafone lacked demand to draw new customers through its services
No tangible rewards to entice users besides number of hours contributed
Limited Experience
Only offered on mobile and only when phone is fully charged
Requires opening app every night, relying on users to actively remember
Difficult to scale due to lack of active user efforts and high friction experience
App launched by Vodafone Foundation and The Garvan Institute of Medical Research in 2015 that allows users to find a cure to cancer by pooling their smartphones’ computing power
“DreamLab aims to create the nation's first "smartphone
supercomputer", and if 100,000 users pooled their phones'
processing capabilities researchers would be able to crunch data
approximately 3,000 times faster than the current rate,” the Institute
said.
DreamLab’s Shortcomings
Appendix: DreamLab
95% of BOINC users are in
North America and Western
Europe
90% of BOINC users are male 35-50 years old with a strong IT background
Median income of BOINC users is highly above average
BOINC Demographics
LEADERBOARD
1 1.5 B (1999)
2 364 M (2017)
3 444 M (2017)
4 80 M (2018)
5 293 M (2001)
6 196 M (2017)
7 478 M (1999)
8 214 M (1999)
9 157 M (2017)
RankTotal Credit
YearJoined
$4.4 M worth of Compute instances
BOINC’s Challenges:
Complex
Privacy
Redundant
Incentive
Lack enterprise functionality due to lack of privacy
Work checked by sending job to two parties, doubling efforts
Volunteer-based fails to gain traction among consumers
Academia focused platform ignores other segments’ UX
Appendix: BOINC
Thanks!