Upload
maria-stylianou
View
730
Download
0
Embed Size (px)
DESCRIPTION
Course: Execution Environments for Distributed Computing 6th Presentation (10-15min): Intelligent Placement of Datacenters for Internet Services Source: http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=5961695
Citation preview
Execution Environments for Distributed Computing
Intelligent Placement of Datacenters for Internet
Services
EEDC
3433
0
European Master in Distributed Computing - EMDC
Homework number: 6
Maria Stylianou [email protected]
2
Outline
● Why is Intelligent Datacenter Placement so important?
● Proposal● Problem Setup
● Parameters● Solving the problem
● Solution Approaches● Evaluation + Selection of Best Approach
● Datacenter Placement Tradeoffs● Conclusions
3
Why is Intelligent Datacenter Placement so important?
Internet Services are hosted in datacenters
4
Why is Intelligent Datacenter Placement so important?
Datacenter Location
Internet Services● Response Time● Costs● Carbon Dioxide
Emissions
5
Why is Intelligent Datacenter Placement so important?
Datacenter Location
Internet Services● Response Time● Costs● Carbon Dioxide
Emissions
Considerations● Proximity to
● Population centers● Power Plants
● Source of electricity● Electricity, land,
water prices● Avg temperatures
6
Proposal
● Framework for the datacenter selection process
● Other Approaches for efficiency
● Build a tool → selecting datacenter locations automatically
7
Problem Setup
An Internet Company wants...
→ Select Locations for datacenters
→ Offer services to Population Centers
while...
→ keeping a minimized overall cost on the datacenter network
→ respecting constraints● network latency● consistency delay● availability
8
Problem Setup
Important Parameters
● Cost
→ Capital & Operational
● Response Time
→ Latency & #servers
● Consistency Delay
→ Latency from a neighbor datacenter
● CO2 emissions
● Service Availability
→ Level of redundancy
Low
High!
9
Problem Setup
Solving the problem● Large # of potential locations to evaluate● Non-linear
● No fast solution● Linear Programming Solvers not applicable
10
Solution Approaches
● Simple Linear Programming (LP0)● Pre-set Linear Programming (LP1)● Brute Force (Brute)● Heuristic based on LP (Heuristic)● Simulated annealing plus LP1 (SA+LP1)● Optimized SA+LP1 (OSA+LP1)
11
Solution Approaches
Evaluation ● Simple Linear Programming (LP0)● Pre-set Linear Programming (LP1)● Brute Force (Brute)● Heuristic based on LP (Heuristic)● Simulated annealing plus LP1 (SA+LP1)● Optimized SA+LP1 (OSA+LP1)
→ Used for comparison
Cannot be used by itself
12
Solution Approaches
Evaluation ● Simple Linear Programming (LP0)● Pre-set Linear Programming (LP1)● Brute Force (Brute)● Heuristic based on LP (Heuristic)● Simulated annealing plus LP1 (SA+LP1)
Optimized SA+LP1 (OSA+LP1)
→ Used for comparison
Cannot be used by itself
13
Datacenter Placement Tradeoffs
● How much does X cost?● Lower Latency: 50ms – best compromise
>70ms → same cost ($7.8M/month)
● Higher Availability: – Tier II datacenters – best option
– Cheaper to build networks with
less redundant datacenters
● Faster Consistency: – in contrast with lower latency
– depends on # locations
14
Datacenter Placement Tradeoffs
● How much does X cost?● Green Datacenter Network:
– Same latency results with an optimal-cost DC
– Less than $100K more expensive
● Chiller-less Datacenter Network:
– latency > 70ms → Cost reduction by 8%
– latency < 70ms → Not possible without chillers
15
Conclusions
● Intelligent Placement mandatory!
→ saves money!
● Linear Programming & Simulated Annealing
Efficient & Accurate Selection Process
Execution Environments for Distributed Computing
Intelligent Placement of Datacenters for Internet
Services
EEDC
3433
0
European Master in Distributed Computing - EMDC
Homework number: 6
Maria Stylianou [email protected]