. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
.
.
. ..
.
.
Intelligent placement of DCs for Internet Services
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and RicardoBianchini
Presentor: Emmanouil Dimogerontakis@{EEDC}
FIB UPC
May 7, 2012
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
.. .1 Overview
.. .2 Framework Description
ParametersThe Optimization Problem & SolutionsExploring the toolEvaluating Solutions
.. .3 Exploring Tradeoffs
.. .4 Future Work
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
.. Datacenters Explosion
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
.. Location
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
.. The problem
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
.. A solution
Provide a framework to automate location selection to satisfy:
Customers Needs
Low Company Costs
Friendliness to environment
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
ParametersThe Optimization Problem & SolutionsExploring the toolEvaluating Solutions
.. .1 Overview
.. .2 Framework Description
ParametersThe Optimization Problem & SolutionsExploring the toolEvaluating Solutions
.. .3 Exploring Tradeoffs
.. .4 Future Work
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
ParametersThe Optimization Problem & SolutionsExploring the toolEvaluating Solutions
.. Costs
COSTS = CAPEX + OPEX
CAPEX i.e.:
land
electricity coverage
network coverage
OPEX i.e.:
maintenance &administration
personnel
electricity & water usage
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
ParametersThe Optimization Problem & SolutionsExploring the toolEvaluating Solutions
.. More
Response Time
Consistency Delay
Availability
CO2 Emisions
All parameters depend directly to location and may contradicteach other.
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
ParametersThe Optimization Problem & SolutionsExploring the toolEvaluating Solutions
.. Problem
Use parameters to formulate a Non-Linear cost model andoptimization problem.
Inputs:
Total max number of servers
Expected average utilization
Users per Server
Redundancy Level
Max network delay &consistency delay
Min availability
CAPEX and OPEX cost ∀locations
Outputs:
Optimal Cost
Max numbers of servers
Number of servers that serveeach population center
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
ParametersThe Optimization Problem & SolutionsExploring the toolEvaluating Solutions
.. Solutions
HARD Problem (non-linear,large number of location).Need for custom solutions.
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)
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
ParametersThe Optimization Problem & SolutionsExploring the toolEvaluating Solutions
.. Input Data
Location-dependent: Datacenter Characteristics:
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
ParametersThe Optimization Problem & SolutionsExploring the toolEvaluating Solutions
.. Characterizing Locations
Figure: Datacenter cost depending on location
Figure: Location characteristicsby Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
ParametersThe Optimization Problem & SolutionsExploring the toolEvaluating Solutions
.. Was it worth it?
Figure: Cost per month distribution in 500 locations of the US
Figure: Cost of a network of datacenters for 60K serversby Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
ParametersThe Optimization Problem & SolutionsExploring the toolEvaluating Solutions
.. Best Algorithm?
OSA+LP1 is the best solution.
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
.. .1 Overview
.. .2 Framework Description
ParametersThe Optimization Problem & SolutionsExploring the toolEvaluating Solutions
.. .3 Exploring Tradeoffs
.. .4 Future Work
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
.. Solutions
Cost vs Latency
Golden ratio: 50 ms
Cost vs Availability
Cheaper if less redundantGolden ratio: Tier II
Cost vs Consistency speed
Consistency and latency are conflicting
Cost vs Environment
for max latencies >70ms: green network $100K moreexpensive per month
Cost vs Chillers
for max latencies >70ms: Avoiding chillers ⇒ ↓ cost 8%
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
.. Solutions
Cost vs Latency
Golden ratio: 50 ms
Cost vs Availability
Cheaper if less redundantGolden ratio: Tier II
Cost vs Consistency speed
Consistency and latency are conflicting
Cost vs Environment
for max latencies >70ms: green network $100K moreexpensive per month
Cost vs Chillers
for max latencies >70ms: Avoiding chillers ⇒ ↓ cost 8%
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
.. Solutions
Cost vs Latency
Golden ratio: 50 ms
Cost vs Availability
Cheaper if less redundantGolden ratio: Tier II
Cost vs Consistency speed
Consistency and latency are conflicting
Cost vs Environment
for max latencies >70ms: green network $100K moreexpensive per month
Cost vs Chillers
for max latencies >70ms: Avoiding chillers ⇒ ↓ cost 8%
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
.. Solutions
Cost vs Latency
Golden ratio: 50 ms
Cost vs Availability
Cheaper if less redundantGolden ratio: Tier II
Cost vs Consistency speed
Consistency and latency are conflicting
Cost vs Environment
for max latencies >70ms: green network $100K moreexpensive per month
Cost vs Chillers
for max latencies >70ms: Avoiding chillers ⇒ ↓ cost 8%
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
.. Solutions
Cost vs Latency
Golden ratio: 50 ms
Cost vs Availability
Cheaper if less redundantGolden ratio: Tier II
Cost vs Consistency speed
Consistency and latency are conflicting
Cost vs Environment
for max latencies >70ms: green network $100K moreexpensive per month
Cost vs Chillers
for max latencies >70ms: Avoiding chillers ⇒ ↓ cost 8%
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
.. .1 Overview
.. .2 Framework Description
ParametersThe Optimization Problem & SolutionsExploring the toolEvaluating Solutions
.. .3 Exploring Tradeoffs
.. .4 Future Work
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services
. . . . . .
OverviewFramework Description
Exploring TradeoffsFuture Work
.. Improvements
How can we improve the framework?
Automatic data collection
Frequency of data updates
Covering other parts of the world
by Inigo Goiri, Kien Le, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presentor: Emmanouil Dimogerontakis @{EEDC}Intelligent placement of DCs for Internet Services