Event-Based Scheduling for Energy-Efficient QoS (eQoS) in Mobile Web Applications
Yuhao Zhu, Matthew Halpern, Vijay Janapa Reddi
Department of Electrical and Computer Engineering, The University of Texas at Austin
HPCA’15
Motivation
Prior art only focused on the trad e-off between raw performance and energy consumption.
◦Ignoring the application QoS characte ristic.
◦Raw performance does not directly cor respond to application QoS.
The Interplay between QoS, Perfor
mance, and Energy.
Contribution
Propose eQoS framework for reasoni ng about the QoS-energy trade-off in mobile Web application.
Propose event-based scheduling.
Propose QPE
◦An eQoS metric that quantifies the tr ade-off between QoS and energy consum ption.
eQoS
Energy-efficient QoS.
A new concept that captures the Qo S-energy trade-off.
Provides “just enough” performan ce to meet users’ QoS expectation s with minimal energy consumption.
◦Imperceptibility
◦Usability
Mobile Web Application
Event-driven
◦Various user interactions, sensor inputs and application internal tasks are transl ated to one or more applications events.
◦Each event is registered with an event ha ndler.
◦FIFO-like event queue.
A software thread continuously monitors the ev ent queue.
dequeues any available event from the head of the queue for processing, one event at a time.
Fundamental Event-level Character istics
Event intensity
◦The frequency of events triggered per second.
Event latency
◦The event execution time.
◦The responsiveness to an event.
Event characteristics
Workload Description
Event Imperceptibility (P
I) and Usability (P
U) Values
Low Event-Intensity, High Event-La tency
◦(P1, PU): (1, 10) s
Low Event-Intensity, Low Event-Lat ency
◦(P1, PU): (50, 100) ms
For web browsing, (P1, PU): (1, 3) ms
High Event-Intensity, Low Event-La tency
◦(P1, PU): (60, 30) FPS
Event-Based Scheduling
Scheduling Unit: event-
handler
Detector
Identifies the P
Iand P
Uvalues fo r an event handler.
◦Based on event latency and event inte nsity information.
◦High latency: latency > 0.8 s
◦High intensity: intensity > 3 times p er second
QoS Monitor
Takes the predictive models, P
Ian d P
Uvalues to determine the archi tecture configuration for executin g a handler.
Monitors event latencies and inten sities on the hardware
◦Adjusts its prediction and scheduling decisions on the fly.
◦Feedback-driven optimizations
Model Constructor
Builds a performance and energy mode l for each event handler.
Performance model:
◦Use the highest and the second-highest f requencies to construct the performance model .
Energy model:
◦Profiling and store in a local power pro file file.
f N
T
memory dependent/ time
Execution
Evaluation
QPE
◦QoS per energy
◦QoS Score: utility function between 0
~1.
(QoSI = 1, QoSU = 0)
n Consumptio Energy
Score
QPE QoS
Experimental Setup
Odroid XU+E development board
◦Samsung Exynos 5410 SoC
◦4 big + 4 little
Android 4.2.2
◦Google’s Chromium Web browser 33.0
Embed all interactions into the be nchmarked applications.
◦Ensuring reproducibility
Model Accuracy
Application:
Paper.js
Compare with Other Schedul ers
Four baseline schedulers:
◦Perf-sched
◦Interactive-sched
◦On-demand-sched
◦Energy-sched
Oracle-sched
◦Has a priori knowledge of all event h andler latencies.
◦Always maximizes the QPE score.
Architecture Configuration Dis
tribution for Imperceptibility
Summary of Comparison
Imperceptibility
◦EBS consumes 0.4% more QoS violation than Perf-sched, but saves on average 41.2% power.
◦EBS achieves 22.9% and 37.9% energy s avings over Ondemand-sched and Intera ctive-sched.
About 0.1% more QoS violation.
◦EBS reduces 72.0% QoS violation compa red to Energy-sched.
Summary of Comparison(Con t.)
Usability
◦EBS achieves 55.4%, 52.9%, and 41.4%
energy savings over Perf-sched, Inter active-sched, and Ondemend-sched, res pectively, with nearly equivalent QoS violations (< 0.1%).
◦Compared to Energy-sched, EBS reduces the QoS violation by about 50%.
Case Study
EDP vs QPE
Big.Little Architectu re
◦ beneficial for eQoS op timizations.
Low-latency, low-intensit y applications (second gr oup) benefit from having a little core.
Applications in the first and third group benefit f rom having a big core.
Conclusion
Propose eQoS, which serves as a gene ral framework for reasoning about th e energy efficiency trade-off in int eractive mobile Web applications.
Demonstrate a working prototype and conduct real hardware and software m easurements.
◦The event-based scheduling optimizing fo r eQoS achieves 41.2% energy saving with only 0.4% of perceptible QoS violations.
