1.3 Energy-Aware Computing
Uploaded by
christse1.3 Energy-Aware Computing
Uploaded by
christseEnergy-Aware Computing
Future of Computing – on the Road to Quantum
Max Plauth
Operating Systems and Middleware Group, HPI
Motivation: Renewable Energy Sources
2 Energy-Aware Computing – Future of Computing | Max Plauth
Motivation: Volatile Energy Market
Fluctuating energy supply leads to
highly variable energy prices
3 Energy-Aware Computing – Future of Computing | Max Plauth
Motivation:
Energy Mgmt. in Data Center Operation
Variable power-target and penalty zones must be
considered in data center operation
4 Energy-Aware Computing – Future of Computing | Max Plauth
Challenge:
How to Influence Power Consumption?
How can power consumption of compute infrastructure be influenced by software?
1. Work another time: Defer/pick jobs with matching energy profile
(popular research in embedded systems,
requires known and deferrable jobs)
2. Work less/more: Use fewer/additional compute resources
(commonly used to scale depending
on operation requirements)
3. Work elsewhere: Use other kinds of compute resources
Focus of our research at the OSM group
Sven Köhler Lukas Wenzel
5 Energy-Aware Computing – Future of Computing | Max Plauth
As general purpose computing power stagnates,
Challenge:
FPGA acceleration can speed up parallel tasks
Use the Right Hardware for the Job
■ The exponential development of computing power described in
Logic Blocks Programmable
Moore‘s law starts to stagnate ■ Fuels development of hardware Interconnect
accelerators
■ Many tasks may benefit from specialized
Hardware: hardware,
NVIDIA GV100 GPU with 84 SMs but custom chip
manufacturing needs high numbers to be profitable
■ Field-Programmable Gate Array: programmable hardware circuit
■ Had a peak in interest more than a decade ago and did not live up to
ParProg 2019
GPUs
the expectations
Max Plauth
Chart 7
■ Hard to integrate in software solutions
RAM/ALU/...
■ Different development paradigm with high learning effort Blocks IO Blocks
CPU GPU FPGA
+ extremely flexible + high throughput + highly optimized
Chart 2
- less specialized / - only suitable for - hard to program
efficient for data-parallel
certain task workloads
6 Energy-Aware Computing – Future of Computing | Max Plauth
Example:
Use the Right Hardware for the Job
■ Power efficiency of different hardware components
for heat dissipation simulation workload (heatmap)
■ Workload uses common compute pattern
□ e.g. applicable in weather forecasts
■ Power draw is collected from various sources
□ Vendor tool (GPU), datasheet (CPU), sim. (FPGA)
■ The FPGA yielded best performance/effieciency
□ Only for the workload at hands
Device Problem Size Cell Type Throughput Power Efficiency
(MCells) (MCells/s) (W) (MCells/Ws)
float
GPU 256 (32bit) 1127,88 60 18,80
float
CPU 256 (32bit) 1435,48 85 16,89
CPU 256 char (8bit) 1420,29 85 16,71
FPGA 1024 char (8bit) 2209,35 9,5 232,56
7 Energy-Aware Computing – Future of Computing | Max Plauth
Challenge:
Measuring Power and Energy
■ There are no standardized programming interfaces available for measuring power draw or
energy consumption of computer hardware
■ Each and any hardware platform provides different means (if any)
■ Example 1: CPU- and GPU-based execution of heatmap simulation on low-power hardware
□ Measured using internal sensors (left), as well as external sensors (right)
□ On this hardware, the GPU is faster / more efficient
□ Internal sensors offer individual readings for each subsystem, but are heavily smoothed
8 Energy-Aware Computing – Future of Computing | Max Plauth
Challenge:
Measuring Power and Energy
■ Example 2: CPU- and GPU-based execution of heatmap simulation on high-performance server
□ Only measurements at the power supply level are available
□ On this machine, the CPU is faster / more energy-efficient
9 Energy-Aware Computing – Future of Computing | Max Plauth
Approach:
Measuring Power and Energy
■ Perf-inspired Energy Profiling Tool (PINPOINT)
□ A cross-platform tool for measuring energy consumption of applications
□ Unifies varying measurement methods available for different platforms / hardware
□ Available on GitHub: https://github.com/osmhpi/pinpoint
$> pinpoint -i 50 -r 10 ./heat 1000 1000 10000 in.csv
[interval: 50ms, repetitions: 10, before: 0ms, after: 0ms]
47436.04 mJ CPU ( +- 4.65% )
8084.43 mJ GPU ( +- 5.22% )
9904.58 mJ SOC ( +- 4.26% )
19315.06 mJ DDR ( +- 3.58% )
95933.73 mJ INPUT ( +- 4.63% )
8.91673945 seconds time elapsed ( +- 5.92% )
10 Energy-Aware Computing – Future of Computing | Max Plauth
Approach:
Trading in Time for Efficiency
■ Mild reductions of the clock frequency may improve energy-efficiency significantly, leading only
to slightly increased execution times.
■ Example: BERT machine learning finetuning workload on GPUs
11 Energy-Aware Computing – Future of Computing | Max Plauth
Outlook
■ With the growing degree of heterogeneity in data centers, choosing the best hardware for each
task will be increasingly important.
□ Applications need to support a larger number of hardware targets.
□ Tooling/infrastructure is needed that makes placement decisions easier.
■ Being able to quantify energy demand of applications is a first but important step that helps
developers to not only to improve for performance, but also for energy-efficiency.
□ Raising awareness of the carbon footprint caused by code is important.
□ Improved energy-tooling enables developers to a developer.
■ For some workloads, energy-efficiency can be improved significantly by sacrificing a small
amount of performance.
□ It remains to be investigated if this approach is applicable to a wirder range of workloads and
hardware platforms.
12 Energy-Aware Computing – Future of Computing | Max Plauth
Make every Joule count!
