Heat-aware Loadbalancing - Is it a thing? Lukas Iffländer, Norbert Schmitt , Andreas Knapp, and Samuel Kounev Chair of Software Engineering – Universität Würzburg November 13, 2020 11th Symposium on Software Performance 2020, 12.–13. November 2020, Leipzig, Germany https://se.informatik.uni-wuerzburg.de/
Motivation Increasing amount of cloud users and services [1] Motivation Goals & Contributions Approach Evaluation Conclusion 2
Motivation Increasing amount of cloud users and services Increasing requirements on computing power by more traffic Increasing computing power needed [1] [2] Motivation Goals & Contributions Approach Evaluation Conclusion 3
Motivation Increasing amount of cloud users and services Increasing requirements on computing power by more traffic Increasing computing power needed [1] Solution 1: More servers? ↯ [2] Motivation Goals & Contributions Approach Evaluation Conclusion 4
Motivation Increasing amount of cloud users and services Increasing requirements on computing power by more traffic Increasing computing power needed [1] Solution 1: More servers? ↯ Solution 2: Better/Full usage of existing computing power [2] make use of unallocated CPU boost power Motivation Goals & Contributions Approach Evaluation Conclusion 5
Goals Foundations > Reasearch Question > Approach > Evaluation > Conclusion Is it possible to exceed permanently and feasibly the clock rate with host relay? 1. Motivation Goals & Contributions Approach Evaluation Conclusion 6
Goals Foundations > Reasearch Question > Approach > Evaluation > Conclusion Is it possible to exceed permanently and feasibly the clock rate with host relay? 1. What is the effect on the computing power compared to 2. normal behavior? Motivation Goals & Contributions Approach Evaluation Conclusion 7
Goals Foundations > Reasearch Question > Approach > Evaluation > Conclusion Is it possible to exceed permanently and feasibly the clock rate with host relay? 1. What is the effect on the computing power compared to 2. normal behavior? What is the effect on the energy consumption compared to normal behavior? 3. Motivation Goals & Contributions Approach Evaluation Conclusion 8
Contributions Foundations > Reasearch Question > Approach > Evaluation > Conclusion 1. A first approach to heat-aware load-balancing Motivation Goals & Contributions Approach Evaluation Conclusion 9
Contributions Foundations > Reasearch Question > Approach > Evaluation > Conclusion 1. A first approach to heat-aware load-balancing 2. A prototype implementation with SDN and an initial evaluation Motivation Goals & Contributions Approach Evaluation Conclusion 10
Approach - Boosting Foundations > Reasearch Question > Approach > Evaluation > Conclusion Intel Turbo Boost: • CPU-Boost ≠ Overclocking in general • CPU Boost respects specifications of hardware • Max. n° of boosted cores and max. frequency is limited by headroom • OS requests more computing power Boost • Example Intel: – Stepwise increasing clock with 133,33 MHz – Time in boost state depends on: Temperature [3] Energy Other common used boost: AMD Turbo Core / AMD Precision Boost Boosts are usually disabled in related work and HPC environments due to unpredictability Motivation Goals & Contributions Approach Evaluation Conclusion 11
Approach Foundations > Reasearch Question > Approach > Evaluation > Conclusion Motivation Goals & Contributions Approach Evaluation Conclusion 12
Approach Foundations > Reasearch Question > Approach > Evaluation > Conclusion Motivation Goals & Contributions Approach Evaluation Conclusion 13
Approach Foundations > Reasearch Question > Approach > Evaluation > Conclusion Motivation Goals & Contributions Approach Evaluation Conclusion 14
Approach Foundations > Reasearch Question > Approach > Evaluation > Conclusion Motivation Goals & Contributions Approach Evaluation Conclusion 15
Approach Foundations > Reasearch Question > Approach > Evaluation > Conclusion Motivation Goals & Contributions Approach Evaluation Conclusion 16
Approach Foundations > Reasearch Question > Approach > Evaluation > Conclusion Motivation Goals & Contributions Approach Evaluation Conclusion 17
Approach Foundations > Reasearch Question > Approach > Evaluation > Conclusion Motivation Goals & Contributions Approach Evaluation Conclusion 18
Approach Foundations > Reasearch Question > Approach > Evaluation > Conclusion Motivation Goals & Contributions Approach Evaluation Conclusion 19
Approach Foundations > Reasearch Question > Approach > Evaluation > Conclusion Motivation Goals & Contributions Approach Evaluation Conclusion 20
Approach Foundations > Reasearch Question > Approach > Evaluation > Conclusion Motivation Goals & Contributions Approach Evaluation Conclusion 21
Approach Foundations > Reasearch Question > Approach > Evaluation > Conclusion Definition: Not max. boosted anymore time to switch Motivation Goals & Contributions Approach Evaluation Conclusion 22
Approach Foundations > Reasearch Question > Approach > Evaluation > Conclusion Assumptions Definition: • The workload puts enough Not max. boosted anymore stress on a server that it will time to switch boost • The server is under stress that the migrated workload will only be adequately serviced when boosted Motivation Goals & Contributions Approach Evaluation Conclusion 23
Prototype Implementation Foundations > Reasearch Question > Approach > Evaluation > Conclusion Client: HTTP Load Generator [4] Motivation Goals & Contributions Approach Evaluation Conclusion 24
Prototype Implementation Foundations > Reasearch Question > Approach > Evaluation > Conclusion Client: HTTP Load Generator Worker: [4] • Workload: Bungee-LU-Servelet [5] • Data collector: Telegraf [6] Motivation Goals & Contributions Approach Evaluation Conclusion 25
Prototype Implementation Foundations > Reasearch Question > Approach > Evaluation > Conclusion Client: HTTP Load Generator Worker: [4] • Workload: Bungee-LU-Servelet [5] • Data collector: Telegraf [6] Monitor [6] • Database: InfluxDB [6] • Dashboard: Chronograf • SDN-Controller framework: Ryu [7] Motivation Goals & Contributions Approach Evaluation Conclusion 26
Prototype Implementation Foundations > Reasearch Question > Approach > Evaluation > Conclusion Client: HTTP Load Generator Worker: [4] • Workload: Bungee-LU-Servelet [5] • Data collector: Telegraf [6] Monitor [6] • Database: InfluxDB [6] • Dashboard: Chronograf • SDN-Controller framework: Ryu [7] Switch: Software-Switch with Open vSwitch [8] Motivation Goals & Contributions Approach Evaluation Conclusion 27
Prototype Implementation Foundations > Reasearch Question > Approach > Evaluation > Conclusion Client: HTTP Load Generator Worker: [4] • Workload: Bungee-LU-Servelet [5] • Data collector: Telegraf [6] Monitor [6] • Database: InfluxDB [6] • Dashboard: Chronograf • SDN-Controller framework: Ryu [7] Switch: Software-Switch with Open vSwitch [8] Testbed with three servers, SDN, and controller Motivation Goals & Contributions Approach Evaluation Conclusion 28
Evaluation Foundations > Reasearch Question > Approach > Evaluation > Conclusion Three evaluated scenarios: A. Highly overbooked A B. Medium overbooked C. Not overbooked Motivation Goals & Contributions Approach Evaluation Conclusion 29
Evaluation Foundations > Reasearch Question > Approach > Evaluation > Conclusion A Three evaluated scenarios: A. Highly overbooked B. Medium overbooked C. Not overbooked B Motivation Goals & Contributions Approach Evaluation Conclusion 30
Evaluation Foundations > Reasearch Question > Approach > Evaluation > Conclusion A Three evaluated scenarios: A. Highly overbooked B. Medium overbooked C. Not overbooked C Motivation Goals & Contributions Approach Evaluation Conclusion 31
Recommend
More recommend