SlideShare a Scribd company logo
Bricklayer: Resource Composition
on the Spot Market
Walter Wong1, Lorenzo Corneo2, Aleksandr Zavodovski1, Pengyuan
Zhou1, Nitinder Mohan1,3, Jussi Kangasharju1
1University of Helsinki, Finland
2Uppsala University, Sweden
3Technical University Munich, Germany
Motivation – AWS Cloud Computing
• Amazon offers a wide range of cloud services, such as computing and
storage, that can be quickly deployed in the Internet
• For computing, there are three subscription options
• Reserved: 1 to 3-year commitment
• On-demand: pay-as-you-go
• Spot instances: variable market price
• AWS Spot Instances
• Up to 90% off compared to the on-demand pricing
• Caveat: no availability guarantee (instances can be shut down with a 2-minute
warning)
Motivation – AWS Cloud & Spot Market
Is it possible to leverage spot instances and provide (almost)
the same availability levels as on-demand instances?
Outline
• AWS Spot Market Analysis
• Bricklayer Proposal
• Evaluation and Experimental Results
• Conclusion
AWS Spot Market Analysis
On-demand price
Market price for spot instance
Main characteristics
• Market price
• Bid Model
• Eviction due to price
AWS has 32 availability zones and 252 instance types resulting in ~8k individual spot markets
AWS Spot Market Analysis
Instance Family Min. ECU (¢/h) Max. ECU (¢/h) Diff. (%)
Compute optimized 0.0026 0.008958 344.54
General purpose 0.003154 0.011854 375.81
Storage optimized 0.014276 0.025212 176.6
Memory optimized 0.00356 0.037692 1058.77
FPGA instances 0.019038 0.021064 110.06
GPU instances 0.022979 0.04365 189.95
Fig. (a) shows that for the same instance family (same hardware), there can be a large ECU price variation
Fig. (b) shows that even with the instance size increase, there is a bottom line for the ECU price.
(a)
(b)
What is the minimum price per computing unit across all instances?
AWS EC2 Compute Unit (ECU) is a benchmark used to compare AWS EC2 instances.
AWS Spot Market Analysis
Fig. (a) and (b) compare the most and least volatile spot instances.
Fig. (c) shows the availability and eviction rates of a sample of instances.
(a) (b) (c)
What are the price volatility of each instance across all availability zones?
What are the instances with lower uptime?
AWS Spot Market Analysis – Lessons Learned
• Spot instances follow the market price rather than a fixed hourly fee
• Eviction happens when the market price of the spot instance goes
over the bid price
• For the same hardware family, we found
• Considerable difference in the ECU pricing
• Difference in the price volatility and instance availability
Bricklayer Goals
• Offer a software tool to analyze the best pricing and availability of
spot instances
• Two optimization models
• Price: searches for the best pricing engines for a given time
• Availability: split the load among different spot markets to reduce eviction
rate
• Leverage the opportunities with the spot instances
• Better ECU pricing with the same hardware
• Lower eviction rates in some instances
Bricklayer – System Design
• Three core metrics
• ECU unit price
• Price volatility (from historical data)
• Availability
• Budget Optimization
• Aims to get the cheapest instances (even repeating the same instance type
multiple times)
• Caveat: bulk eviction (all instances are swiped away at the same time)
• Availability Optimization
• Aims to get different instance types across different availability zones to
reduce evictions
Bricklayer – Architecture
Linear Optimizer implements Google Operation Research library
All components expose a REST-API to be consumed independently (microservices)
Evaluation Setup
• Applications Categories
• Batch-type applications (fault tolerant, non-time critical)
• Highly available applications
• Data
• AWS spot pricing data (March to June, 2019)
• Contains 32 availability zones and 252 instance types
• Instance Sizes
• Four popular sizes
• 16 vCPU (60 ECUs), 32 vCPU (131 ECUs), 64 vCPU (262 ECUs) and 96 vCPU
(347 ECUs)
Results – Budget Optimized Selection
Comparison between selecting a pre-defined instance type, e.g., m5.4xlarge, and
using Bricklayer to provide a resource set with the same amount of vCPUs
Results: 51% to 88% cheaper compared to Spot instances and 83% to 95% cheaper compared to on-demand prices
16 vCPU 32 vCPU 64 vCPU 96 vCPU
Results – Selecting different spot markets
Spread across multiple distinct spot markets using Bricklayer
Results: increase of 24% going from 1 to 2 spot markets. Further increase 3% per spot market.
Results – Total cost comparison vs. number of
distinct spot markets
Total cost over a 90-day period.
Bricklayer selected the least volatile instances (which have not been evicted due to pricing)
Key Takeaways
• For the budget optimized, we are always cheaper than the regular
spot instances
• Reduction between 51% to 88% off compared to spot instances
• Reduction between 83% to 95% off compared to on-demand instances
• Even adding more more spot markets to increase the availability,
Bricklayer still provides a cheaper solution compared to regular
instances
• The provided spot instances did not experience any eviction due to
pricing
Conclusion
• Bricklayer offers options to reduce the cloud costs by trading off spot
instance’s availability vs price
• Bricklayer checks three metrics to offer the best resource set
optimized for cost or availability
• ECU unit price
• Price volatility
• Eviction rate due to pricing
• The experimental results show that Bricklayer provides same
availability but at lower costs compared to regular instances
Questions?
• Contact Info
• Walter Wong
• University of Helsinki
• Email: walter.wong@helsinki.fi

More Related Content

Similar to Bricklayer: Resource Composition on the Spot Market

Introduction to Amazon EC2 Spot
Introduction to Amazon EC2 Spot Introduction to Amazon EC2 Spot
Introduction to Amazon EC2 Spot
Amazon Web Services
 
IncQuery-D: Distributed Incremental Model Queries over the Cloud: Engineerin...
IncQuery-D: Distributed Incremental Model Queries over the Cloud: Engineerin...IncQuery-D: Distributed Incremental Model Queries over the Cloud: Engineerin...
IncQuery-D: Distributed Incremental Model Queries over the Cloud: Engineerin...
Daniel Varro
 
Cost optimization at scale toronto v3
Cost optimization at scale toronto v3Cost optimization at scale toronto v3
Cost optimization at scale toronto v3
Amazon Web Services
 
From TCO to Optimization at Scale - Pop-up Loft TLV 2017
From TCO to Optimization at Scale - Pop-up Loft TLV 2017From TCO to Optimization at Scale - Pop-up Loft TLV 2017
From TCO to Optimization at Scale - Pop-up Loft TLV 2017
Amazon Web Services
 
Kubernetes: Reducing Infrastructure Cost & Complexity
Kubernetes: Reducing Infrastructure Cost & ComplexityKubernetes: Reducing Infrastructure Cost & Complexity
Kubernetes: Reducing Infrastructure Cost & Complexity
DevOps.com
 
Cc
CcCc
Amazon EC2 is a web service provided by Amazon Web Services (AWS) that offers...
Amazon EC2 is a web service provided by Amazon Web Services (AWS) that offers...Amazon EC2 is a web service provided by Amazon Web Services (AWS) that offers...
Amazon EC2 is a web service provided by Amazon Web Services (AWS) that offers...
NoorUlHaq47
 
Kinney j aws
Kinney j awsKinney j aws
Kinney j aws
souvikbiswas67
 
AWS re:Invent 2016: Beeswax: Building a Real-Time Streaming Data Platform on ...
AWS re:Invent 2016: Beeswax: Building a Real-Time Streaming Data Platform on ...AWS re:Invent 2016: Beeswax: Building a Real-Time Streaming Data Platform on ...
AWS re:Invent 2016: Beeswax: Building a Real-Time Streaming Data Platform on ...
Amazon Web Services
 
AWS APAC Webinar Week - Maintaining Performance & Availability While Lowering...
AWS APAC Webinar Week - Maintaining Performance & Availability While Lowering...AWS APAC Webinar Week - Maintaining Performance & Availability While Lowering...
AWS APAC Webinar Week - Maintaining Performance & Availability While Lowering...
Amazon Web Services
 
Prediction as a service with ensemble model in SparkML and Python ScikitLearn
Prediction as a service with ensemble model in SparkML and Python ScikitLearnPrediction as a service with ensemble model in SparkML and Python ScikitLearn
Prediction as a service with ensemble model in SparkML and Python ScikitLearn
Josef A. Habdank
 
Webinar: How to Size a Lab
Webinar: How to Size a LabWebinar: How to Size a Lab
Webinar: How to Size a Lab
Lizzy Guido (she/her)
 
AWS Summit 2013 | Auckland - Optimizing Your AWS Applications and Usage to Re...
AWS Summit 2013 | Auckland - Optimizing Your AWS Applications and Usage to Re...AWS Summit 2013 | Auckland - Optimizing Your AWS Applications and Usage to Re...
AWS Summit 2013 | Auckland - Optimizing Your AWS Applications and Usage to Re...
Amazon Web Services
 
Cost Optimisation on AWS
Cost Optimisation on AWSCost Optimisation on AWS
Cost Optimisation on AWS
Amazon Web Services
 
AWS Cost Optimization
AWS Cost OptimizationAWS Cost Optimization
AWS Cost Optimization
Miles Ward
 
Optimizing Costs and Efficiency of AWS Services
Optimizing Costs and Efficiency of AWS Services Optimizing Costs and Efficiency of AWS Services
Optimizing Costs and Efficiency of AWS Services
Amazon Web Services
 
Penny Pinching at Scale
Penny Pinching at ScalePenny Pinching at Scale
Penny Pinching at Scale
Amazon Web Services
 
Introduction to Amazon EC2
Introduction to Amazon EC2Introduction to Amazon EC2
Introduction to Amazon EC2
Amazon Web Services
 
AWS Webcast - Webinar Series for State and Local Government #2: Discover the ...
AWS Webcast - Webinar Series for State and Local Government #2: Discover the ...AWS Webcast - Webinar Series for State and Local Government #2: Discover the ...
AWS Webcast - Webinar Series for State and Local Government #2: Discover the ...
Amazon Web Services
 
Augmenting Machine Learning with Databricks Labs AutoML Toolkit
Augmenting Machine Learning with Databricks Labs AutoML ToolkitAugmenting Machine Learning with Databricks Labs AutoML Toolkit
Augmenting Machine Learning with Databricks Labs AutoML Toolkit
Databricks
 

Similar to Bricklayer: Resource Composition on the Spot Market (20)

Introduction to Amazon EC2 Spot
Introduction to Amazon EC2 Spot Introduction to Amazon EC2 Spot
Introduction to Amazon EC2 Spot
 
IncQuery-D: Distributed Incremental Model Queries over the Cloud: Engineerin...
IncQuery-D: Distributed Incremental Model Queries over the Cloud: Engineerin...IncQuery-D: Distributed Incremental Model Queries over the Cloud: Engineerin...
IncQuery-D: Distributed Incremental Model Queries over the Cloud: Engineerin...
 
Cost optimization at scale toronto v3
Cost optimization at scale toronto v3Cost optimization at scale toronto v3
Cost optimization at scale toronto v3
 
From TCO to Optimization at Scale - Pop-up Loft TLV 2017
From TCO to Optimization at Scale - Pop-up Loft TLV 2017From TCO to Optimization at Scale - Pop-up Loft TLV 2017
From TCO to Optimization at Scale - Pop-up Loft TLV 2017
 
Kubernetes: Reducing Infrastructure Cost & Complexity
Kubernetes: Reducing Infrastructure Cost & ComplexityKubernetes: Reducing Infrastructure Cost & Complexity
Kubernetes: Reducing Infrastructure Cost & Complexity
 
Cc
CcCc
Cc
 
Amazon EC2 is a web service provided by Amazon Web Services (AWS) that offers...
Amazon EC2 is a web service provided by Amazon Web Services (AWS) that offers...Amazon EC2 is a web service provided by Amazon Web Services (AWS) that offers...
Amazon EC2 is a web service provided by Amazon Web Services (AWS) that offers...
 
Kinney j aws
Kinney j awsKinney j aws
Kinney j aws
 
AWS re:Invent 2016: Beeswax: Building a Real-Time Streaming Data Platform on ...
AWS re:Invent 2016: Beeswax: Building a Real-Time Streaming Data Platform on ...AWS re:Invent 2016: Beeswax: Building a Real-Time Streaming Data Platform on ...
AWS re:Invent 2016: Beeswax: Building a Real-Time Streaming Data Platform on ...
 
AWS APAC Webinar Week - Maintaining Performance & Availability While Lowering...
AWS APAC Webinar Week - Maintaining Performance & Availability While Lowering...AWS APAC Webinar Week - Maintaining Performance & Availability While Lowering...
AWS APAC Webinar Week - Maintaining Performance & Availability While Lowering...
 
Prediction as a service with ensemble model in SparkML and Python ScikitLearn
Prediction as a service with ensemble model in SparkML and Python ScikitLearnPrediction as a service with ensemble model in SparkML and Python ScikitLearn
Prediction as a service with ensemble model in SparkML and Python ScikitLearn
 
Webinar: How to Size a Lab
Webinar: How to Size a LabWebinar: How to Size a Lab
Webinar: How to Size a Lab
 
AWS Summit 2013 | Auckland - Optimizing Your AWS Applications and Usage to Re...
AWS Summit 2013 | Auckland - Optimizing Your AWS Applications and Usage to Re...AWS Summit 2013 | Auckland - Optimizing Your AWS Applications and Usage to Re...
AWS Summit 2013 | Auckland - Optimizing Your AWS Applications and Usage to Re...
 
Cost Optimisation on AWS
Cost Optimisation on AWSCost Optimisation on AWS
Cost Optimisation on AWS
 
AWS Cost Optimization
AWS Cost OptimizationAWS Cost Optimization
AWS Cost Optimization
 
Optimizing Costs and Efficiency of AWS Services
Optimizing Costs and Efficiency of AWS Services Optimizing Costs and Efficiency of AWS Services
Optimizing Costs and Efficiency of AWS Services
 
Penny Pinching at Scale
Penny Pinching at ScalePenny Pinching at Scale
Penny Pinching at Scale
 
Introduction to Amazon EC2
Introduction to Amazon EC2Introduction to Amazon EC2
Introduction to Amazon EC2
 
AWS Webcast - Webinar Series for State and Local Government #2: Discover the ...
AWS Webcast - Webinar Series for State and Local Government #2: Discover the ...AWS Webcast - Webinar Series for State and Local Government #2: Discover the ...
AWS Webcast - Webinar Series for State and Local Government #2: Discover the ...
 
Augmenting Machine Learning with Databricks Labs AutoML Toolkit
Augmenting Machine Learning with Databricks Labs AutoML ToolkitAugmenting Machine Learning with Databricks Labs AutoML Toolkit
Augmenting Machine Learning with Databricks Labs AutoML Toolkit
 

Recently uploaded

国外证书(Lincoln毕业证)新西兰林肯大学毕业证成绩单不能毕业办理
国外证书(Lincoln毕业证)新西兰林肯大学毕业证成绩单不能毕业办理国外证书(Lincoln毕业证)新西兰林肯大学毕业证成绩单不能毕业办理
国外证书(Lincoln毕业证)新西兰林肯大学毕业证成绩单不能毕业办理
zoowe
 
Understanding User Behavior with Google Analytics.pdf
Understanding User Behavior with Google Analytics.pdfUnderstanding User Behavior with Google Analytics.pdf
Understanding User Behavior with Google Analytics.pdf
SEO Article Boost
 
7 Best Cloud Hosting Services to Try Out in 2024
7 Best Cloud Hosting Services to Try Out in 20247 Best Cloud Hosting Services to Try Out in 2024
7 Best Cloud Hosting Services to Try Out in 2024
Danica Gill
 
manuaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaal
manuaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaalmanuaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaal
manuaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaal
wolfsoftcompanyco
 
Ready to Unlock the Power of Blockchain!
Ready to Unlock the Power of Blockchain!Ready to Unlock the Power of Blockchain!
Ready to Unlock the Power of Blockchain!
Toptal Tech
 
成绩单ps(UST毕业证)圣托马斯大学毕业证成绩单快速办理
成绩单ps(UST毕业证)圣托马斯大学毕业证成绩单快速办理成绩单ps(UST毕业证)圣托马斯大学毕业证成绩单快速办理
成绩单ps(UST毕业证)圣托马斯大学毕业证成绩单快速办理
ysasp1
 
存档可查的(USC毕业证)南加利福尼亚大学毕业证成绩单制做办理
存档可查的(USC毕业证)南加利福尼亚大学毕业证成绩单制做办理存档可查的(USC毕业证)南加利福尼亚大学毕业证成绩单制做办理
存档可查的(USC毕业证)南加利福尼亚大学毕业证成绩单制做办理
fovkoyb
 
Meet up Milano 14 _ Axpo Italia_ Migration from Mule3 (On-prem) to.pdf
Meet up Milano 14 _ Axpo Italia_ Migration from Mule3 (On-prem) to.pdfMeet up Milano 14 _ Axpo Italia_ Migration from Mule3 (On-prem) to.pdf
Meet up Milano 14 _ Axpo Italia_ Migration from Mule3 (On-prem) to.pdf
Florence Consulting
 
办理毕业证(NYU毕业证)纽约大学毕业证成绩单官方原版办理
办理毕业证(NYU毕业证)纽约大学毕业证成绩单官方原版办理办理毕业证(NYU毕业证)纽约大学毕业证成绩单官方原版办理
办理毕业证(NYU毕业证)纽约大学毕业证成绩单官方原版办理
uehowe
 
Gen Z and the marketplaces - let's translate their needs
Gen Z and the marketplaces - let's translate their needsGen Z and the marketplaces - let's translate their needs
Gen Z and the marketplaces - let's translate their needs
Laura Szabó
 
制作原版1:1(Monash毕业证)莫纳什大学毕业证成绩单办理假
制作原版1:1(Monash毕业证)莫纳什大学毕业证成绩单办理假制作原版1:1(Monash毕业证)莫纳什大学毕业证成绩单办理假
制作原版1:1(Monash毕业证)莫纳什大学毕业证成绩单办理假
ukwwuq
 
Discover the benefits of outsourcing SEO to India
Discover the benefits of outsourcing SEO to IndiaDiscover the benefits of outsourcing SEO to India
Discover the benefits of outsourcing SEO to India
davidjhones387
 
Design Thinking NETFLIX using all techniques.pptx
Design Thinking NETFLIX using all techniques.pptxDesign Thinking NETFLIX using all techniques.pptx
Design Thinking NETFLIX using all techniques.pptx
saathvikreddy2003
 
留学挂科(UofM毕业证)明尼苏达大学毕业证成绩单复刻办理
留学挂科(UofM毕业证)明尼苏达大学毕业证成绩单复刻办理留学挂科(UofM毕业证)明尼苏达大学毕业证成绩单复刻办理
留学挂科(UofM毕业证)明尼苏达大学毕业证成绩单复刻办理
uehowe
 
假文凭国外(Adelaide毕业证)澳大利亚国立大学毕业证成绩单办理
假文凭国外(Adelaide毕业证)澳大利亚国立大学毕业证成绩单办理假文凭国外(Adelaide毕业证)澳大利亚国立大学毕业证成绩单办理
假文凭国外(Adelaide毕业证)澳大利亚国立大学毕业证成绩单办理
cuobya
 
学位认证网(DU毕业证)迪肯大学毕业证成绩单一比一原版制作
学位认证网(DU毕业证)迪肯大学毕业证成绩单一比一原版制作学位认证网(DU毕业证)迪肯大学毕业证成绩单一比一原版制作
学位认证网(DU毕业证)迪肯大学毕业证成绩单一比一原版制作
zyfovom
 
Should Repositories Participate in the Fediverse?
Should Repositories Participate in the Fediverse?Should Repositories Participate in the Fediverse?
Should Repositories Participate in the Fediverse?
Paul Walk
 
办理毕业证(UPenn毕业证)宾夕法尼亚大学毕业证成绩单快速办理
办理毕业证(UPenn毕业证)宾夕法尼亚大学毕业证成绩单快速办理办理毕业证(UPenn毕业证)宾夕法尼亚大学毕业证成绩单快速办理
办理毕业证(UPenn毕业证)宾夕法尼亚大学毕业证成绩单快速办理
uehowe
 
重新申请毕业证书(RMIT毕业证)皇家墨尔本理工大学毕业证成绩单精仿办理
重新申请毕业证书(RMIT毕业证)皇家墨尔本理工大学毕业证成绩单精仿办理重新申请毕业证书(RMIT毕业证)皇家墨尔本理工大学毕业证成绩单精仿办理
重新申请毕业证书(RMIT毕业证)皇家墨尔本理工大学毕业证成绩单精仿办理
vmemo1
 
Explore-Insanony: Watch Instagram Stories Secretly
Explore-Insanony: Watch Instagram Stories SecretlyExplore-Insanony: Watch Instagram Stories Secretly
Explore-Insanony: Watch Instagram Stories Secretly
Trending Blogers
 

Recently uploaded (20)

国外证书(Lincoln毕业证)新西兰林肯大学毕业证成绩单不能毕业办理
国外证书(Lincoln毕业证)新西兰林肯大学毕业证成绩单不能毕业办理国外证书(Lincoln毕业证)新西兰林肯大学毕业证成绩单不能毕业办理
国外证书(Lincoln毕业证)新西兰林肯大学毕业证成绩单不能毕业办理
 
Understanding User Behavior with Google Analytics.pdf
Understanding User Behavior with Google Analytics.pdfUnderstanding User Behavior with Google Analytics.pdf
Understanding User Behavior with Google Analytics.pdf
 
7 Best Cloud Hosting Services to Try Out in 2024
7 Best Cloud Hosting Services to Try Out in 20247 Best Cloud Hosting Services to Try Out in 2024
7 Best Cloud Hosting Services to Try Out in 2024
 
manuaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaal
manuaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaalmanuaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaal
manuaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaal
 
Ready to Unlock the Power of Blockchain!
Ready to Unlock the Power of Blockchain!Ready to Unlock the Power of Blockchain!
Ready to Unlock the Power of Blockchain!
 
成绩单ps(UST毕业证)圣托马斯大学毕业证成绩单快速办理
成绩单ps(UST毕业证)圣托马斯大学毕业证成绩单快速办理成绩单ps(UST毕业证)圣托马斯大学毕业证成绩单快速办理
成绩单ps(UST毕业证)圣托马斯大学毕业证成绩单快速办理
 
存档可查的(USC毕业证)南加利福尼亚大学毕业证成绩单制做办理
存档可查的(USC毕业证)南加利福尼亚大学毕业证成绩单制做办理存档可查的(USC毕业证)南加利福尼亚大学毕业证成绩单制做办理
存档可查的(USC毕业证)南加利福尼亚大学毕业证成绩单制做办理
 
Meet up Milano 14 _ Axpo Italia_ Migration from Mule3 (On-prem) to.pdf
Meet up Milano 14 _ Axpo Italia_ Migration from Mule3 (On-prem) to.pdfMeet up Milano 14 _ Axpo Italia_ Migration from Mule3 (On-prem) to.pdf
Meet up Milano 14 _ Axpo Italia_ Migration from Mule3 (On-prem) to.pdf
 
办理毕业证(NYU毕业证)纽约大学毕业证成绩单官方原版办理
办理毕业证(NYU毕业证)纽约大学毕业证成绩单官方原版办理办理毕业证(NYU毕业证)纽约大学毕业证成绩单官方原版办理
办理毕业证(NYU毕业证)纽约大学毕业证成绩单官方原版办理
 
Gen Z and the marketplaces - let's translate their needs
Gen Z and the marketplaces - let's translate their needsGen Z and the marketplaces - let's translate their needs
Gen Z and the marketplaces - let's translate their needs
 
制作原版1:1(Monash毕业证)莫纳什大学毕业证成绩单办理假
制作原版1:1(Monash毕业证)莫纳什大学毕业证成绩单办理假制作原版1:1(Monash毕业证)莫纳什大学毕业证成绩单办理假
制作原版1:1(Monash毕业证)莫纳什大学毕业证成绩单办理假
 
Discover the benefits of outsourcing SEO to India
Discover the benefits of outsourcing SEO to IndiaDiscover the benefits of outsourcing SEO to India
Discover the benefits of outsourcing SEO to India
 
Design Thinking NETFLIX using all techniques.pptx
Design Thinking NETFLIX using all techniques.pptxDesign Thinking NETFLIX using all techniques.pptx
Design Thinking NETFLIX using all techniques.pptx
 
留学挂科(UofM毕业证)明尼苏达大学毕业证成绩单复刻办理
留学挂科(UofM毕业证)明尼苏达大学毕业证成绩单复刻办理留学挂科(UofM毕业证)明尼苏达大学毕业证成绩单复刻办理
留学挂科(UofM毕业证)明尼苏达大学毕业证成绩单复刻办理
 
假文凭国外(Adelaide毕业证)澳大利亚国立大学毕业证成绩单办理
假文凭国外(Adelaide毕业证)澳大利亚国立大学毕业证成绩单办理假文凭国外(Adelaide毕业证)澳大利亚国立大学毕业证成绩单办理
假文凭国外(Adelaide毕业证)澳大利亚国立大学毕业证成绩单办理
 
学位认证网(DU毕业证)迪肯大学毕业证成绩单一比一原版制作
学位认证网(DU毕业证)迪肯大学毕业证成绩单一比一原版制作学位认证网(DU毕业证)迪肯大学毕业证成绩单一比一原版制作
学位认证网(DU毕业证)迪肯大学毕业证成绩单一比一原版制作
 
Should Repositories Participate in the Fediverse?
Should Repositories Participate in the Fediverse?Should Repositories Participate in the Fediverse?
Should Repositories Participate in the Fediverse?
 
办理毕业证(UPenn毕业证)宾夕法尼亚大学毕业证成绩单快速办理
办理毕业证(UPenn毕业证)宾夕法尼亚大学毕业证成绩单快速办理办理毕业证(UPenn毕业证)宾夕法尼亚大学毕业证成绩单快速办理
办理毕业证(UPenn毕业证)宾夕法尼亚大学毕业证成绩单快速办理
 
重新申请毕业证书(RMIT毕业证)皇家墨尔本理工大学毕业证成绩单精仿办理
重新申请毕业证书(RMIT毕业证)皇家墨尔本理工大学毕业证成绩单精仿办理重新申请毕业证书(RMIT毕业证)皇家墨尔本理工大学毕业证成绩单精仿办理
重新申请毕业证书(RMIT毕业证)皇家墨尔本理工大学毕业证成绩单精仿办理
 
Explore-Insanony: Watch Instagram Stories Secretly
Explore-Insanony: Watch Instagram Stories SecretlyExplore-Insanony: Watch Instagram Stories Secretly
Explore-Insanony: Watch Instagram Stories Secretly
 

Bricklayer: Resource Composition on the Spot Market

  • 1. Bricklayer: Resource Composition on the Spot Market Walter Wong1, Lorenzo Corneo2, Aleksandr Zavodovski1, Pengyuan Zhou1, Nitinder Mohan1,3, Jussi Kangasharju1 1University of Helsinki, Finland 2Uppsala University, Sweden 3Technical University Munich, Germany
  • 2. Motivation – AWS Cloud Computing • Amazon offers a wide range of cloud services, such as computing and storage, that can be quickly deployed in the Internet • For computing, there are three subscription options • Reserved: 1 to 3-year commitment • On-demand: pay-as-you-go • Spot instances: variable market price • AWS Spot Instances • Up to 90% off compared to the on-demand pricing • Caveat: no availability guarantee (instances can be shut down with a 2-minute warning)
  • 3. Motivation – AWS Cloud & Spot Market Is it possible to leverage spot instances and provide (almost) the same availability levels as on-demand instances?
  • 4. Outline • AWS Spot Market Analysis • Bricklayer Proposal • Evaluation and Experimental Results • Conclusion
  • 5. AWS Spot Market Analysis On-demand price Market price for spot instance Main characteristics • Market price • Bid Model • Eviction due to price AWS has 32 availability zones and 252 instance types resulting in ~8k individual spot markets
  • 6. AWS Spot Market Analysis Instance Family Min. ECU (¢/h) Max. ECU (¢/h) Diff. (%) Compute optimized 0.0026 0.008958 344.54 General purpose 0.003154 0.011854 375.81 Storage optimized 0.014276 0.025212 176.6 Memory optimized 0.00356 0.037692 1058.77 FPGA instances 0.019038 0.021064 110.06 GPU instances 0.022979 0.04365 189.95 Fig. (a) shows that for the same instance family (same hardware), there can be a large ECU price variation Fig. (b) shows that even with the instance size increase, there is a bottom line for the ECU price. (a) (b) What is the minimum price per computing unit across all instances? AWS EC2 Compute Unit (ECU) is a benchmark used to compare AWS EC2 instances.
  • 7. AWS Spot Market Analysis Fig. (a) and (b) compare the most and least volatile spot instances. Fig. (c) shows the availability and eviction rates of a sample of instances. (a) (b) (c) What are the price volatility of each instance across all availability zones? What are the instances with lower uptime?
  • 8. AWS Spot Market Analysis – Lessons Learned • Spot instances follow the market price rather than a fixed hourly fee • Eviction happens when the market price of the spot instance goes over the bid price • For the same hardware family, we found • Considerable difference in the ECU pricing • Difference in the price volatility and instance availability
  • 9. Bricklayer Goals • Offer a software tool to analyze the best pricing and availability of spot instances • Two optimization models • Price: searches for the best pricing engines for a given time • Availability: split the load among different spot markets to reduce eviction rate • Leverage the opportunities with the spot instances • Better ECU pricing with the same hardware • Lower eviction rates in some instances
  • 10. Bricklayer – System Design • Three core metrics • ECU unit price • Price volatility (from historical data) • Availability • Budget Optimization • Aims to get the cheapest instances (even repeating the same instance type multiple times) • Caveat: bulk eviction (all instances are swiped away at the same time) • Availability Optimization • Aims to get different instance types across different availability zones to reduce evictions
  • 11. Bricklayer – Architecture Linear Optimizer implements Google Operation Research library All components expose a REST-API to be consumed independently (microservices)
  • 12. Evaluation Setup • Applications Categories • Batch-type applications (fault tolerant, non-time critical) • Highly available applications • Data • AWS spot pricing data (March to June, 2019) • Contains 32 availability zones and 252 instance types • Instance Sizes • Four popular sizes • 16 vCPU (60 ECUs), 32 vCPU (131 ECUs), 64 vCPU (262 ECUs) and 96 vCPU (347 ECUs)
  • 13. Results – Budget Optimized Selection Comparison between selecting a pre-defined instance type, e.g., m5.4xlarge, and using Bricklayer to provide a resource set with the same amount of vCPUs Results: 51% to 88% cheaper compared to Spot instances and 83% to 95% cheaper compared to on-demand prices 16 vCPU 32 vCPU 64 vCPU 96 vCPU
  • 14. Results – Selecting different spot markets Spread across multiple distinct spot markets using Bricklayer Results: increase of 24% going from 1 to 2 spot markets. Further increase 3% per spot market.
  • 15. Results – Total cost comparison vs. number of distinct spot markets Total cost over a 90-day period. Bricklayer selected the least volatile instances (which have not been evicted due to pricing)
  • 16. Key Takeaways • For the budget optimized, we are always cheaper than the regular spot instances • Reduction between 51% to 88% off compared to spot instances • Reduction between 83% to 95% off compared to on-demand instances • Even adding more more spot markets to increase the availability, Bricklayer still provides a cheaper solution compared to regular instances • The provided spot instances did not experience any eviction due to pricing
  • 17. Conclusion • Bricklayer offers options to reduce the cloud costs by trading off spot instance’s availability vs price • Bricklayer checks three metrics to offer the best resource set optimized for cost or availability • ECU unit price • Price volatility • Eviction rate due to pricing • The experimental results show that Bricklayer provides same availability but at lower costs compared to regular instances
  • 18. Questions? • Contact Info • Walter Wong • University of Helsinki • Email: walter.wong@helsinki.fi