Geographical load balancing
Jonathan Lukkien 31 mei 2013
Overview
Introduction
Geographical Load Balancing
The Model
Algorithms & Experiments
Extensions
Conclusion
Problem
Introduction1. Data centers are big players in energy consumption
2. How can we get data centers to use less (fossil) fuel?
Solutions
Introduction1. Make data centers more energy effective
2. Geographical Load Balancing
Current practice
Geographical Load BalancingGeographical Load Balancing is used to reduce costs. Paradox: cost reduction by increasing energy consumption!
2.1 The Model
General Idea
Geographical Load BalancingConsidering |N | data centers and |J | sources from queries to these data centers. Model the energy cost per data center for handling some query. Model the lost revenue “cost” incurred by delay on response per data center for some query. Combine these to model complete cost of query handling.
Introducing variables, energy
cost
Geographical Load Balancingξi(t) = gi(t, mi(t), λi(t))
with
ξi(t) denoting the energy cost for data center i.
mi(t) denoting the amount of servers active at data center i.
λi(t) denoting the load at data center i.
Introducing variables, delay
cost
Geographical Load Balancingδi(t) =Pj∈Jλij(t)r(fi(mi(t), λi(t)) + dij(t))
with
δi(t) denoting the delay cost for data center i.
λij(t) denoting the amount of work sent to data center i from
source j.
r(d) denoting the revenue lost for a job given delay d. fi(mi, λi) denoting the amount of delay caused by queueing.
dij(t) denoting the amount of delay caused by the network
The GLB
Geographical Load Balancing minPT t=1 P i∈N(ξi(t) + δi(t)) s.t. P i∈Nλij(t) = Lj(t) ∀j ∈ J λij(t) ≥ 0 ∀i ∈ N, ∀j ∈ J 0 ≤ mi(t) ≤ Mi ∀i ∈ N mi(t) ∈ N ∀i ∈ NThis is annoying to solve.
Easier GLB
Geographical Load BalancingWe drop the integrality constraint and instead we round up when we find a solution.
Now we have an easier problem to solve. But still problems remain, some suggestions?
Some sidenotes
Geographical Load Balancing1. An optimal solution which has non-sparse routing can be transformed to one that has sparse routing.
2. Rerouting of queries is necessary for this to work, research has been done on this subject.
3. Adapting capacity of data centers should be possible on the fly, also research being done.
4. Estimation of propagation delay should be reasonably accurate.
5. We do not want to solve it centrally!
A distributed approach
Geographical Load BalancingTwo algorithms proposed
1. Algorithm 1 has every source solving the LP with some partial information
2. Algorithm 2 has every source solving steepest gradient with additional smart things
Case Studies
Geographical Load BalancingTwo case studies were done:
1. Case study 1 considers data center perspective
Case study 1: setup
Geographical Load BalancingStudy used historical data to model input data. 14 data centers were used.
Service time per query is set to 1. Static pricing of electricity is used.
Case study 1:
benchmarks
Geographical Load BalancingThree alternative strategies are also tested to provide some context
1. Baseline 1: does not consider energy pricing when minimizing cost.
2. Baseline 2: does not consider network delay when minimizing cost.
3. Baseline 3: considers neither network delay nor energy pricing.
Case study 1: results
Geographical Load BalancingThe optimal algorithm suggested by the authors is best.
Case study 2: setup
Geographical Load BalancingBuilding on the setup of case study 1 three things are added:
1. A model for availability of renewable energy sources.
2. The pricing for renewable energy.
Case study 2: results
Geographical Load BalancingGeographical load balancing has a trend to follow renewables. However, some renewables are more beneficial than others to follow.
1. Taking in to account switching on and off of servers.
2. Providing better estimates for delays in the system.
3. Incorporate more renewables in the model.
We have seen the idea of geographical load balancing. Distributed algorithms are in place to reduce energy costs. Dynamic pricing is needed to make sure renewables can be in corporated.
Discussion
ConclusionIn practice not widely used for following renewables... yet! How long will it take?