CSR-Small-Energy-Aware Redundancy Management
1
EEL Lab Computer Science CSR:Small: Energy-Aware Redundancy Management Target Applications RAPID System Overview RSDG (Redundant Services Dependence Graph) Motivation 1. Applications are composed of sets of collaborating services (producer- consumer, resource sharing) 2. Each individual service may provide discrete levels of quality vs. energy (resource) tradeoffs. 3. Nothing is free! User wants best overall application outcome within a given energy budget. 4. Need to deal with failure and uncertainty . 5. Need to support user customizable quality metrics. 2 energy consumption mission value 3 types of redundancy: - approximation - implementation - replication Service Selection and Productivity Profile RSDG service selection problem: - Services with mission values are considered (user) critical - For each critical service, select a single service level and implementation; do the same for each service the selected service depends on (transitive dependent set) - Sum of mission values of critical services is maximal under given energy constrains; energy is minimized for this maximal mission value Summary and Conclusion A new RAPID framework explicitly models and represents redundancies through a graph representation (RSDG). This allows efficient search space explorations. The system expert provides the basic structure of the application using his/her domain knowledge. An automatic tool derives the hard cost metrics (energy) through online and offline training. The application user customizes the service preferences and specifies an overall energy budget. RAPID then adaptively balances the services configurations across the entire system to produce highest user productivity with lowest cost within the user-specific energy budget, even in the presence of failures and uncertainty. This is accomplished by reconfiguring the system in response to such unexpected events. Future work includes merging RSDGs and energy budgets of simultaneously executing applications. Solution for energy = 40 and example mission values: 11 Solution for energy = 31 and example mission values: 9 Solution for energy = 29 and example mission values: 7 Solution for energy = 12 and example mission values: 5 Optimal solution balances service qualities across the entire application; Solution seems sometimes non-intuitive. Solution is NP-complete; Proof: Reduction from 3SAT. Use of 0-1 integer programming formulation and Gurobi to solve the problem optimally. Simulations can provide feedback during tradeoff space exploration via the Productivity Profile. RSDG Underwater Glider (Slocum) Multiple sensors and dependencies between services (dep. edges not shown) RSDG NavApp on Smartphone Navigation application on Android that balances multiple services to meet the energy requirement. PI: Ulrich Kremer NSF-CSR #1617551 Services Redundancy levels Dependencies Mission values (service priori5es) Energy budget & mission 5me Environmental configura5on Dual%weighted full RSDG with service selec6on Dynamic service selec6on problem > maximize produc5vity > minimize energy Op5mal and heuris5c solu5ons User System Expert System RSDG Automa6c Tool Power dissipa5on of > service levels > communica5on > transi5ons System RSDG with power annota6ons Dual%weighted Full RSDG Dynamic reconfigura6on triggered by > failure condi5ons > changed energy budget > environmental events > fixed 5me interval s 4b s 4a 5 S S S 1 S 2 3 4a s s 4b 10 5 energy cost S S S 1 S 2 3 5 2 2 15 15 2 2 15 3 8 12 4b s 4a s 10 5 energy cost mission value S S S 1 S 3 2 5 6 2 3 5 2 2 15 15 2 3 15 8 12 2 4b s 4a s 5 selection with: mission value energy cost energy = 29 mission value = 7 S S 3 S 2 1 S 10 5 6 2 3 5 2 2 15 15 15 2 2 3 8 12 RSDG Service Robot Expressiveness of RSDG to encode route planning and service selection problem. Productivity profiles of RAPID vs. heuristic with different user preferences: top: 3:1.5 (2:1), bottom 1:1 Productivity profiles of RAPID vs. heuristic with different user preferences: top: 6:6:6 (1:1:1), bottom 10:1:1 Experiments Exp1: Highest service levels w/o reconfiguration. Exp2: RAPID with screen preference and stable energy consumption. Exp3: RAPID with screen preference and unexpected energy drain. Exp4: RAPID with localization preference and unexpected energy drain. Exp5: Lowest service levels w/o reconfiguration but w/ unexpected energy drain. Exp1 Exp2 Exp3 Exp4 Exp5 NavApp: RAPID runs on LG Nexus 5 (Android), Quad-core, 2.3 GHz; Gurobi runs on remote server; receives 0-1 problem from phone, and returns result; power measured by Qualcomm’s Trepn tool. Overheads: 4 ms solution, 0.3 ms reconfiguration; 0.09% energy. Service Robot: RAPID runs on ARM- based NVIDIA Jetson TK1 system; Gurobi runs remotely as in NavApp; power measured by on-board Pololu ACS711EX current sensor. Overheads: negligible due to small problem size (>1 ms solve + reconfig); energy dominated by motors.
Transcript of CSR-Small-Energy-Aware Redundancy Management
EEL LabComputer Science
CSR:Small: Energy-Aware Redundancy Management
Target Applications RAPID System Overview
RSDG (Redundant Services Dependence Graph)
Motivation
1. Applications are composed of sets of collaborating services (producer- consumer, resource sharing)
2. Each individual service may provide discrete levels of quality vs. energy (resource) tradeoffs.
3. Nothing is free! User wants best overall application outcome within a given energy budget.
4. Need to deal with failure and uncertainty. 5. Need to support user customizable quality metrics.
2
energy consumption
mission value
3 types of redundancy: - approximation - implementation - replication
Service Selection and Productivity Profile
RSDG service selection problem: - Services with mission values are considered (user) critical - For each critical service, select a single service level and implementation; do the same for each service the selected service depends on (transitive dependent set) - Sum of mission values of critical services is maximal under given energy constrains; energy is minimized for this maximal mission value
Summary and ConclusionA new RAPID framework explicitly models and represents redundancies through a graph representation (RSDG). This allows efficient search space explorations. The system expert provides the basic structure of the application using his/her domain knowledge. An automatic tool derives the hard cost metrics (energy) through online and offline training. The application user customizes the service preferences and specifies an overall energy budget. RAPID then adaptively balances the services configurations across the entire system to produce highest user productivity with lowest cost within the user-specific energy budget, even in the presence of failures and uncertainty. This is accomplished by reconfiguring the system in response to such unexpected events. Future work includes merging RSDGs and energy budgets of simultaneously executing applications.
Solution for energy = 40 and example mission values: 11
Solution for energy = 31 and example mission values: 9
Solution for energy = 29 and example mission values: 7
Solution for energy = 12 and example mission values: 5
Optimal solution balances service qualities across the entire application; Solution seems sometimes non-intuitive. Solution is NP-complete; Proof: Reduction from 3SAT. Use of 0-1 integer programming formulation and Gurobi to solve the problem optimally. Simulations can provide feedback during tradeoff space exploration via the Productivity Profile.
RSDG Underwater Glider (Slocum) Multiple sensors and dependencies between services (dep. edges not shown)
RSDG NavApp on Smartphone Navigation application on Android that balances multiple services to meet the energy requirement.
PI: Ulrich Kremer NSF-CSR #1617551
Services(Redundancy(levels(Dependencies(
Mission(values((service(priori5es)(Energy(budget(&(mission(5me(Environmental(configura5on(
Dual%weighted-full-RSDG-with-service-selec6on-
Dynamic-service-selec6on-problem-(((((>(maximize(produc5vity((((((>(minimize(energy(
Op5mal(and(heuris5c((solu5ons(
User-System-Expert-
System-RSDG-
Automa6c-Tool-Power(dissipa5on(of((((((>(service(levels((((((>(communica5on((((((>(transi5ons((
System-RSDG--with-power-annota6ons-
Dual%weighted-Full-RSDG--
Dynamic-reconfigura6on-triggered-by-(((((>(failure(condi5ons((((((>(changed(energy(budget((((((>(environmental(events((((((>(fixed(5me(interval(
s 4bs4a
5
S
SS1
S
2
3
4as s4b
10
5
energy cost
S
SS1
S
2
3
522
1515
2
215
3
812
4bs4as
10
5
energy costmission value
S
SS1
S3
2
5 6
23 522
1515
2
3
15
812
2
4bs4as
5
selection with:
mission value
energy cost
energy = 29mission value = 7 S
S3
S 21 S
105 6
23 522
15
1515
2
2
3
812
RSDG Service Robot Expressiveness of RSDG to encode route planning and service selection problem.
Productivity profiles of RAPID vs. heuristic with different user preferences: top: 3:1.5 (2:1), bottom 1:1
Productivity profiles of RAPID vs. heuristic with different user preferences: top: 6:6:6 (1:1:1), bottom 10:1:1
ExperimentsExp1: Highest service levels w/o reconfiguration. Exp2: RAPID with screen preference and stable energy consumption. Exp3: RAPID with screen preference and unexpected energy drain. Exp4: RAPID with localization preference and unexpected energy drain. Exp5: Lowest service levels w/o reconfiguration but w/ unexpected energy drain.
Exp1
Exp2
Exp3
Exp4
Exp5
NavApp: RAPID runs on LG Nexus 5 (Android), Quad-core, 2.3 GHz; Gurobi runs on remote server; receives 0-1 problem from phone, and returns result; power measured by Qualcomm’s Trepn tool. Overheads: 4 ms solution, 0.3 ms reconfiguration; 0.09% energy.
Service Robot: RAPID runs on ARM-based NVIDIA Jetson TK1 system; Gurobi runs remotely as in NavApp; power measured by on-board Pololu ACS711EX current sensor. Overheads: negligible due to small problem size (>1 ms solve + reconfig); energy dominated by motors.
