Empya : Saving Energy in the Face of Varying Workloads Christopher - - PowerPoint PPT Presentation

▶
empya saving energy in the face of varying workloads
SMART_READER_LITE
LIVE PREVIEW

Empya : Saving Energy in the Face of Varying Workloads Christopher - - PowerPoint PPT Presentation

Aug 12, 2023 691 likes •941 views

Empya : Saving Energy in the Face of Varying Workloads Christopher Eibel , Thao-Nguyen Do, Robert Meiner, and Tobias Distler System Software Group Friedrich-Alexander-Universitt Erlangen-Nrnberg (FAU) The 2018 IEEE International

slide-1
SLIDE 1

Empya: Saving Energy in the Face of Varying Workloads

Christopher Eibel, Thao-Nguyen Do, Robert Meißner, and Tobias Distler

System Software Group Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU)

The 2018 IEEE International Conference on Cloud Engineering (IC2E ’18) April 20, 2018

slide-2
SLIDE 2

Introduction & Motivation

Execution platforms vs. energy demand in data centers

Programming and execution platforms are generally not energy aware Dynamic applications in data centers are faced with varying workloads Resources are often statically assigned ➔ Consequence: Lots of energy is being wasted

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Introduction & Motivation 2

slide-3
SLIDE 3

Introduction & Motivation

Execution platforms vs. energy demand in data centers

Programming and execution platforms are generally not energy aware Dynamic applications in data centers are faced with varying workloads Resources are often statically assigned ➔ Consequence: Lots of energy is being wasted

Example application: key–value store

Receives and processes user requests with basic operations (e.g., get(key)) Programmers may choose between two configuration options:

➔ Staticenergy: Lower performance when load is high ➔ Staticperf : Wasting energy when load is low

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Introduction & Motivation 2

slide-4
SLIDE 4

Introduction & Motivation

Execution platforms vs. energy demand in data centers

Programming and execution platforms are generally not energy aware Dynamic applications in data centers are faced with varying workloads Resources are often statically assigned ➔ Consequence: Lots of energy is being wasted

Example application: key–value store

Receives and processes user requests with basic operations (e.g., get(key)) Programmers may choose between two configuration options:

➔ Staticenergy: Lower performance when load is high ➔ Staticperf : Wasting energy when load is low

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Introduction & Motivation 2

5 10 15 20 Staticperf Staticenergy Power demand at 70 kOps/s Power demand [W] 100 200 300 400 Staticperf Staticenergy Maximum throughput Throughput [kOps/s]

slide-5
SLIDE 5

Introduction & Motivation

Execution platforms vs. energy demand in data centers

Programming and execution platforms are generally not energy aware Dynamic applications in data centers are faced with varying workloads Resources are often statically assigned ➔ Consequence: Lots of energy is being wasted

Example application: key–value store

Receives and processes user requests with basic operations (e.g., get(key)) Programmers may choose between two configuration options:

➔ Staticenergy: Lower performance when load is high ➔ Staticperf : Wasting energy when load is low

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Introduction & Motivation 2

5 10 15 20 Staticperf Staticenergy Power demand at 70 kOps/s Power demand [W] 100 200 300 400 Staticperf Staticenergy Maximum throughput Throughput [kOps/s]

Goals

Control the energy–performance tradeoff Propose platform that

❶ frees programmers from taking care of energy optimizations ❷ uses available techniques at hardware and software level ❸ adapts dynamically to varying workloads

slide-6
SLIDE 6

General Approach

EMPYA: energy-aware middleware platform for dynamic applications Key design principles for EMPYA

❶ Energy-efficiency awareness

Avoid high CPU utilization because of disproportionate power-to-performance ratio Not necessarily select configuration with full resource allocation

❷ Multi-level awareness

Exploit available techniques at multiple levels Coordinate techniques: Best energy efficiency with respect to required performance

❸ Energy awareness

Integrated regulator making energy-aware reconfigurations No additional services

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Approach 3

slide-7
SLIDE 7

EMPYA – Exploiting Techniques at Different Levels

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Approach 4

Platform Level OS Level Hardware Level Energy Regulator

slide-8
SLIDE 8

EMPYA – Exploiting Techniques at Different Levels

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Approach 4

➔ Vary #threads ➔ Mapping of application components to threads

Platform Level OS Level Hardware Level Energy Regulator

slide-9
SLIDE 9

EMPYA – Exploiting Techniques at Different Levels

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Approach 4

➔ Vary #threads ➔ Mapping of application components to threads ➔ Vary #(un)active cores ➔ Mapping of application threads to active cores

Platform Level OS Level Hardware Level Energy Regulator

slide-10
SLIDE 10

EMPYA – Exploiting Techniques at Different Levels

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Approach 4

➔ Vary #threads ➔ Mapping of application components to threads ➔ Vary #(un)active cores ➔ Mapping of application threads to active cores ➔ Vary upper power limits ➔ Instruct hardware to enforce them

Platform Level OS Level Hardware Level Energy Regulator

slide-11
SLIDE 11

EMPYA – Exploiting Techniques at Different Levels

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Approach 4

➔ Vary #threads ➔ Mapping of application components to threads ➔ Vary #(un)active cores ➔ Mapping of application threads to active cores ➔ Vary upper power limits ➔ Instruct hardware to enforce them

Platform Level OS Level Hardware Level Energy Regulator

slide-12
SLIDE 12

EMPYA – Exploiting Techniques at Different Levels

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Approach 5

Platform Level OS Level Hardware Level Energy Regulator A0 A1 A2

Thread Pool0 Thread Pool1

C0 C1 C2 C3

Active Active Inactive Active

Power Control Energy Accounting

slide-13
SLIDE 13

EMPYA – Exploiting Techniques at Different Levels

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Approach 5

❶ Actors

Each actor maintains its own state Communication via message passing Actor is independent

  • f executing thread

➔ Implementation: Akka toolkit

Platform Level OS Level Hardware Level Energy Regulator A0 A1 A2

Thread Pool0 Thread Pool1

C0 C1 C2 C3

Active Active Inactive Active

Power Control Energy Accounting

slide-14
SLIDE 14

EMPYA – Exploiting Techniques at Different Levels

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Approach 5

❶ Actors

Each actor maintains its own state Communication via message passing Actor is independent

  • f executing thread

➔ Implementation: Akka toolkit

❷ Power limiting

Running average power limit (RAPL) Originally developed for power limiting (e.g., temperature issues) Enables power and energy measurements Power capping very powerful for reducing the energy demand

Platform Level OS Level Hardware Level Energy Regulator A0 A1 A2

Thread Pool0 Thread Pool1

C0 C1 C2 C3

Active Active Inactive Active

Power Control Energy Accounting

slide-15
SLIDE 15

EMPYA – Energy Regulator

Self-adapting system with continuous feedback loop

Monitor application performance Emit dynamic HW/SW reconfigurations

Energy-profile database

Configuration characteristics Workload-specific power values

Energy policies

Primary performance goal (e.g., throughput) Secondary performance goal (e.g., latency)

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Approach 6

Platform Level OS Level Hardware Level

Energy Regulator

Observer Configurator Control Unit Profiles

Energy Policy

getValues() reconfigure()

slide-16
SLIDE 16

EMPYA – Energy Regulator

Self-adapting system with continuous feedback loop

Monitor application performance Emit dynamic HW/SW reconfigurations

Energy-profile database

Configuration characteristics Workload-specific power values

Energy policies

Primary performance goal (e.g., throughput) Secondary performance goal (e.g., latency)

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Approach 6

Platform Level OS Level Hardware Level

Energy Regulator

Observer Configurator Control Unit Profiles

Energy Policy

getValues() reconfigure()

ID Configuration Performance Power #Threads #Cores Cap Throughput Latency usage α 24 8 None 390.5 kOps/s 0.42 ms 51.2 W 70.4 kOps/s 0.37 ms 19.3 W λ 12 6 22 W 224.8 kOps/s 0.62 ms 22.0 W 50.5 kOps/s 0.25 ms 15.3 W ω 1 1 10 W 20.6 kOps/s 0.22 ms 10.0 W 15.1 kOps/s 0.21 ms 9.7 W

slide-17
SLIDE 17

EMPYA – Energy Regulator

Self-adapting system with continuous feedback loop

Monitor application performance Emit dynamic HW/SW reconfigurations

Energy-profile database

Configuration characteristics Workload-specific power values

Energy policies

Primary performance goal (e.g., throughput) Secondary performance goal (e.g., latency)

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Approach 6

Platform Level OS Level Hardware Level

Energy Regulator

Observer Configurator Control Unit Profiles

Energy Policy

getValues() reconfigure() energy policy { application = key-value-store; throughput_min_ops_per_sec = 10k; throughput_priority = pri; latency_max_msec = 0.5; latency_priority = sec; }

ID Configuration Performance Power #Threads #Cores Cap Throughput Latency usage α 24 8 None 390.5 kOps/s 0.42 ms 51.2 W 70.4 kOps/s 0.37 ms 19.3 W λ 12 6 22 W 224.8 kOps/s 0.62 ms 22.0 W 50.5 kOps/s 0.25 ms 15.3 W ω 1 1 10 W 20.6 kOps/s 0.22 ms 10.0 W 15.1 kOps/s 0.21 ms 9.7 W

slide-18
SLIDE 18

Evaluation – Evaluation Setup

Hardware

Client and server machines, switched 1 Gbps Ethernet Intel Xeon E3-1245 v3 & Xeon E3-1275 v5 processors 8 cores with Hyper-Threading enabled, 3.40 GHz Speed Step and TurboBoost enabled

Application classes

Use case A: Key–value store with mixed operations (get, set, exists) Use case B: MapReduce running single, different jobs

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Evaluation 7

Xeon Client machine Key1 Values1 Key2 Values2 Keyn Valuesn

. . . . . .

Data Map Shuffle Reduce Output

slide-19
SLIDE 19

Evaluation – Use Case A: Key–Value Store

Staticperf vs. EMPYA Throughput as primary performance goal

60 120 180 240 300 360 420 100 200 300 400 Time [s] Throughput [kOps/s] Staticperf EMPYA 60 120 180 240 300 360 420 10 20 30 40 50 60 Time [s] Power [W] Staticperf EMPYA

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Evaluation 8

slide-20
SLIDE 20

Evaluation – Use Case A: Key–Value Store

Staticperf vs. EMPYAlatency Throughput as primary and latency as secondary performance goal

60 120 180 240 300 360 420 0.5 1.0 1.5 2.0 Time [s] Latency [ms] Staticperf EMPYAdefault EMPYAlatency 60 120 180 240 300 360 420 10 20 30 40 50 60 Time [s] Power [W] Staticperf EMPYAdefault EMPYAlatency

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Evaluation 8

slide-21
SLIDE 21

Evaluation – Use Case B: MapReduce

Staticenergy/Staticperf vs. EMPYA Performance goal: Specifying maximum execution-time penalties

1 Energy

Staticperf Staticenergy Empya5% Empya15% Empya30%

1 2 Runtime

1507.7 J 517.5 J 752.9 J 594.2 J 529.6 J 2982.5 J 1456.6 J 2124.9 J 1827.7 J 1815.1 J 2402.4 J 887.1 J 1537.8 J 1346.1 J 889.7 J 2527.7 J 843.9 J 1271.6 J 1244.3 J 947.1 J 1967.3 J 724.8 J 1328.0 J 923.4 J 817.9 J 643.0 J 242.4 J 423.3 J 397.9 J 367.5 J 1415.8 J 432.8 J 511.4 J 481.3 J 443.2 J 64.9 s 84.3 s 66.2 s 72.9 s 83.5 s 102.1 s 196.0 s 109.8 s 117.5 s 123.7 s 62.2 s 119.0 s 63.3 s 64.3 s 77.2 s 67.6 s 113.6 s 70.6 s 74.8 s 92.8 s 60.1 s 98.1 s 61.5 s 68.4 s 75.1 s 15.3 s 32.5 s 15.9 s 16.0 s 20.4 s 41.9 s 59.1 s 44.2 s 45.2 s 51.7 s

sort kmeans wc topn mean join

  • verlap

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Evaluation 9

slide-22
SLIDE 22

Conclusion

EMPYA

Self-adaptive middleware platform enforcing HW and SW reconfigurations Exploiting actors and operating-system functionality Power capping as an effective power- and energy-reduction measure

Key–value store: Up to 34 % less power demand MapReduce: Energy savings of 22–64 %

Future and ongoing work

Making decisions in a distributed manner for multiple machines Carefully increasing the configuration space → heterogeneity

  • C. Eibel, C. Gulden, W. Schröder-Preikschat, and T. Distler

Strome: Energy-Aware Data-Stream Processing

In Proceedings of the 18th IFIP International Conference on Distributed Applications and Interoperable Systems (DAIS 2018), 2018.

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Conclusion 10

slide-23
SLIDE 23

Conclusion

EMPYA

Self-adaptive middleware platform enforcing HW and SW reconfigurations Exploiting actors and operating-system functionality Power capping as an effective power- and energy-reduction measure

Key–value store: Up to 34 % less power demand MapReduce: Energy savings of 22–64 %

Future and ongoing work

Making decisions in a distributed manner for multiple machines Carefully increasing the configuration space → heterogeneity

  • C. Eibel, C. Gulden, W. Schröder-Preikschat, and T. Distler

Strome: Energy-Aware Data-Stream Processing

In Proceedings of the 18th IFIP International Conference on Distributed Applications and Interoperable Systems (DAIS 2018), 2018.

IC2E ’18 Empya: Saving Energy in the Face of Varying Workloads Conclusion 10

Thank you for your attention.

Questions?