scaling energy adaptive applications for sustainable...
TRANSCRIPT
Scaling Energy Adaptive Applications for Sustainable Profitability
Fabien Hermenier, Giovanni Giuliani, Andre Milani, Sophie Demassey
Let existing and new data centres become energy adaptive
Being adapted to the requests of a Smart City Energy Management Authority
Adapting the power consumption to the availability of renewable energy
How to reconcile competing objectives ?
ClientEnergy authorityregulate
energy usage
An economical approach to avoid overcommitment
best performance
forecasts
energyproviders
sustainable objectiveEasc
application descriptor
elected working modes
smart cityauthority
Easc...Carverweb service
videotranscoder
energy adaptiveapplications
Orchestrates EASC for profitable sustainability
Optimises over 24-hrs time window (96 time-slots of 15 minutes)
Called every 15 minutes to accommodate uncertainty
[ucc 2015]
The energy providers
Per time-slot data or forecasts for the next 24 hours
% issues from renewable sources price
capacity
day-to-day periods
Objective
Regulate energy usage through contracts
Incentive
At least X% renewable, Power budget, …
Penalty function (flat, linear, composite)
Smart city
energy authority
EASC characterisation
daily-based deferrable workloads
Service Level Objectives
flat or linear pricing policies
slot-based non-deferrable workloads
performance level power consumption
transition cost actuator
The working modes
discretise elastic applications states
manual or automatic calibration
3 use cases inside DC4Cities
Webservice, video transcoding,
e-health
For every time-slot
find 1 working mode per EASC
using available energy (viable dispatch)
+ transition costs + various Energy Authority policies + Instant and cumulative objectives + generic pricing policies
96 NP-hard bin-packing problems to solve
The underlying problem
I dont want a heuristic full of corner cases=
the right model for the right problem
deterministic composition high-level constraints
An Open-Source java library for constraint programming
time5 6 7 8 9 10 11 12
-3 -5 -3
An automaton to model each EASC life-cycle
Counters to accumulate daily incomes
Modelling - TLDR;
∞
Energy to use is dispatched among the sources
Energy authority pricing policy over the cumulative daily usage
Working mode
Reconciliations Trading energy, performance and energy authority conformance to maximise the daily running costs
“
Evaluating Carver
real deployments 3 testbeds running production softwares
CSUC @ Barcelona
Video transcoding openNebula stack
Trento @ Italy
e-health Openstack
Web service Mixed EASCs Bare metal
HP Lab @ Milan
20 HP moonshot cartridges in 2 chassis no virtualisation layer
20 Watts peak
8m2 of solar panels
4 typical days from a 1 y. data collect
Testbed
0
200
400
600
16/01 17/01 18/01 19/01 20/01Date
Powe
r (W
)
0
25
50
75
100
16/01 17/01 18/01 19/01 20/01Date
rene
wabl
e po
wer (
%)
65% renewable target 100€ / pp.
un-pure energy market price
The grid
The green
Energy authority expectations
1000
1500
2000
2500
3000
23:00 05:00 11:00 17:00 23:00Time
Request/s
e-learning courses for Entrepreneurs
instant SLO 6 working modes
penalty from a step function
Injectors mimic the production workload
production software stack with 3 EASCs
Web servicecumulative SLOs 3 working modes
penalty from a linear function
E-learning, G-learning
Production mode
Cache built at midnight
Lowest suitable working mode for the WebSite
Most efficient working mode when renewable energy is here
Lowest otherwise
Perf
Green
CarverSustainable profitability
VS
0
100
200
300
400
500
16/01 17/01 18/01 19/01 20/01Date
Watts
applicationE−learningG−indexingWebsite
Perf
Green Carver
Batch activities follow the sun Green is binary
Ignores energy availability
Workload affinity: human work during days
0
200
400
16/01 17/01 18/01 19/01 20/01Date
Watts
applicationE−learning
G−indexing
Website
0
200
400
16/01 17/01 18/01 19/01 20/01Date
Watts
applicationE−learning
G−indexing
Website
17/01/15 18/01/15 19/01/15 20/01/15
64.2 65.667.7
55.657.9 58.5
53.956.7 57.3
45.7 46.8 47.4
40
50
60
70
perf
Carver
green pe
rf
Carver
green pe
rf
Carver
green pe
rf
Carver
green
rene
wabl
e %
17/01/15 18/01/15 19/01/15 20/01/151999
1466
1999 1999
1352
1933 1999
1305
1936 1999
1072
1955
0
500
1000
1500
2000
perf
green
Carver pe
rfgre
enCarv
er perf
green
Carver pe
rfgre
enCarv
er
SLO
pen
alty
(eur
os)
Carver sticked to the green threshold
Carver sticked to the SLO
Customer incomes
Carver do not over-commit
max achievement
% renewable
Green neglects the clients
Perf neglects the energy authority
Carver trades
17/01/15 18/01/15 19/01/15 20/01/15
0
1000
2000
perf
green
Carver pe
rfgre
enCarv
er perf
green
Carver pe
rfgre
enCarv
er
Run
ning
cos
t (eu
ros)
expenseenergySLOSMA
17/01/15 18/01/15 19/01/15 20/01/15
0
1000
2000
perf
green
Carver pe
rfgre
enCarv
er perf
green
Carver pe
rfgre
enCarv
er
Run
ning
cos
t (eu
ros)
expenseenergySLOSMA
The balance is still slightly pure performance oriented
nothing to do to please the energy authority (natural workload affinity)
@Day 1
@Day 4 Only minor trading possibilities
every price <<< pricing policies to avoid bankruptcy
Economically speakingEnergy price does not play a role
small data centres human resources / software support dominates
17/01/15 18/01/15 19/01/15 20/01/15
0
1000
2000
perf
green
Carver pe
rfgre
enCarv
er perf
green
Carver pe
rfgre
enCarv
er
Run
ning
cos
t (eu
ros)
expenseenergySLOSMA
Lesson learned
being generic is costlyCP composability helps but data makes the problem
static analysis for stronger models
require favorable hardware and software
variable working modes energy-efficient hardware
a multifacet tool PV array sizer cost modelling prospective deployment
Carver is looking for sun and profit
A flexible solving algorithm to cope with the problem variability
A multi-faceted tool
Conciliation possibilities validated on industrial testbeds
Sustainable profitability to motivate energy transition
http://www.dc4cities.eudeliverables, scientific publications, trial results,
software repositories