SLIDE 1
University of Minnesota
Clouds
Cloud: Hides details of actual service deployment
Nebulas: Using Distributed Voluntary Resources to Build Clouds - - PowerPoint PPT Presentation
Nebulas: Using Distributed Voluntary Resources to Build Clouds - - PowerPoint PPT Presentation
Nebulas: Using Distributed Voluntary Resources to Build Clouds Abhishek Chandra and Jon Weissman Department of Computer Science University of Minnesota University of Minnesota Clouds Cloud: Hides details of actual service deployment from
SLIDE 2
SLIDE 3
University of Minnesota
Current Cloud Model
Cloud: Hides details of actual service deployment
from users
Users
SLIDE 4
University of Minnesota
Current Cloud Model
Largely centralized (or small degree of
distribution)
Pay-as-you-go model Strong guarantees Question: Are there services that do not need/
fit this cloud model?
SLIDE 5
University of Minnesota
Class 1: “Experimental” Services
Experimental deployment for:
Debugging, viability, requirement estimation
SLIDE 6
University of Minnesota
Class 1: “Experimental” Services
SOSP 07 USENIX 09 OSDI 08
SOSP 2009 Experimental deployment for:
Debugging, viability, requirement estimation
SLIDE 7
University of Minnesota
Class 2: Dispersed-Data-Intensive Services
Data is geographically distributed
Costly, inefficient to move to central location
SLIDE 8
University of Minnesota
Class 2: Dispersed-Data-Intensive Services
blog1 blog2 blog3
Data is geographically distributed
Costly, inefficient to move to central location
SLIDE 9
University of Minnesota
Class 3: Shared “Public” Services
Tour
- f
Paris Personal application offered as free service
User-demand driven, scale-up/scale-down needed
SLIDE 10
University of Minnesota
Class 3: Shared “Public” Services
Tour
- f
Paris Personal application offered as free service
User-demand driven, scale-up/scale-down needed
SLIDE 11
University of Minnesota
Common Service Characteristics
Elastic resource consumption
Scale up/down based on demand
Geographical data/user distribution
Execution dependent on location of data/user
Low/no cost
Do not want to pay for resources
Weak performance/robustness requirements
Some failures may be ok
SLIDE 12
University of Minnesota
Users
Cloud
Cloud: Hides details of actual service deployment
from users
SLIDE 13
University of Minnesota
Nebula
Decentralized, less-managed cloud
Dispersed storage/compute resources No/low user cost
Users
SLIDE 14
University of Minnesota
Building Nebulas
Idea: Use distributed voluntary
resources
Resources donated by end-users ala @home, P2P systems
SLIDE 15
University of Minnesota
Why Voluntary Resources?
Scalability: Large number of resources available
SETI@Home: Over 2.2 million computers contributing
~510 TFlops of compute power
Kazaa: Over 3.5 million users
Low cost:
Minimal deployment, management costs [Kondo09]: 2 orders of magnitude difference in EC2
- vs. SETI@home resources/$
Dispersion: Geographically distributed
Users can be located worldwide
SLIDE 16
University of Minnesota
How is Nebula different from @home?
Cloud-oriented services impose new requirements
Requirement Nebula @home Collective performance High None Locality/Context- awareness High Low Statefulness High/medium Low
SLIDE 17
University of Minnesota
17
Challenges
Heterogeneity
Different nodes have different CPU speeds, network
bandwidth, loads
Resource dispersion
Data sources and compute resources may be widely
distributed
Unreliability
Node/link failures, high churn
SLIDE 18
University of Minnesota
Handling Heterogeneity
Heterogeneity-aware resource selection and allocation
Allows better collective performance Trivedi et al. [IJHPCA06]: Fit tasks to node capability
Heterogeneity-aware allocation reduces execution time
SLIDE 19
University of Minnesota
Handling Data Dependence
Find compute nodes and data sources with high
accessibility to each other
Kim et al. [UM-TR08]: Use passive accessibility estimation
Data accessibility-based selection improves download time
SLIDE 20
University of Minnesota
Handling Failures
Replication, state-maintenance Sonnek et al. [TPDS07]: Reliability-aware dynamic replication 20
Dynamic replication improves performance, reliability
SLIDE 21
University of Minnesota
Other Issues/Challenges
Incentivizing Nebulas
Market economy, bartering, auctions How to prevent cheating/freeloading?
Deployment tools/APIs/client support
Virtualization, Middleware?
Privacy/security issues
How to secure systems and applications? We think: Nebulas not suitable for privacy-
sensitive services
SLIDE 22
University of Minnesota
Summary
Current Cloud models:
Well-provisioned, well-managed, centralized
Some service classes:
Need loose performance, low/no cost, distributed
data-intensive
Nebula: Distributed, less-managed clouds
Use voluntary resources