High-Performance and Energy- Efficient Mobile Web Browsing on Big/Little Systems Yuhao Zhu Vijay Janapa Reddi Trinity Research Group The University of Texas at Austin
www.cnn.com 12 Mobile systems demand high performance! ARM Cortex A9 @ 1.2 GHz Webpages are becoming computationally intensive Increasing Computational Intensity
www.cnn.com Can we deploy a mobile system purely driven by performance objectives? 12 Mobile systems demand high performance! ARM Cortex A9 @ 1.2 GHz Webpages are becoming computationally intensive Increasing Computational Intensity
No. Mobile devices are battery-constrained! www.cnn.com Can we deploy a mobile system purely driven by performance objectives? 12 Mobile systems demand high performance! ARM Cortex A9 @ 1.2 GHz Webpages are becoming computationally intensive Increasing Computational Intensity
Executive Summary 13 Challenge: How to design the system architecture that guarantees both high performance and energy efficiency? Alternatives: Single big/little core; symmetric designs; asymmetric designs;
Challenge: How to design the system architecture that guarantees both high performance and energy efficiency? Alternatives: Single big/little core; symmetric designs; asymmetric designs; Big/Little systems • Different microarchitectures (Big, OoO + little, in-order) • Different operating points (DVFS) 14 Executive Summary
Challenge: How to design the system architecture that guarantees both high performance and energy efficiency? Alternatives: Single big/little core; symmetric designs; asymmetric designs; Big/Little systems • Different microarchitectures (Big, OoO + little, in-order) • Different operating points (DVFS) Key insight: Webpages have different characteristics that lead to load time and energy consumption variance 14 Executive Summary
Challenge: How to design the system architecture that guarantees both high performance and energy efficiency? Alternatives: Single big/little core; symmetric designs; asymmetric designs; Big/Little systems • Different microarchitectures (Big, OoO + little, in-order) • Different operating points (DVFS) Key insight: Webpages have different characteristics that lead to load time and energy consumption variance 14 Solution: Predict configuration and schedule webpages accordingly Executive Summary
Software Setup • We studied the Gecko rendering engine in Firefox • Excluded boot-strap and shut-down effects • Disabled browser cache • Hottest 5,000 webpages from www.alexa.com • Downloaded and mapped to the memory 15 Independent of the particular browser
Hardware Setup Big core: ARM Cortex A9: OoO with 4 issue (e.g. in Tegra 3-based tablets) PandaBoard ES Rev B1, 45nm DVFS: 350 MHz, 0.83 V DVFS: 700 MHz, 1.01 V DVFS: 920 MHz, 1.11 V DVFS: 1.2 GHz, 1.27 V 16
Hardware Setup Big core: ARM Cortex A9: OoO with 4 issue (e.g. in Tegra 3-based tablets) PandaBoard ES Rev B1, 45nm DVFS: 350 MHz, 0.83 V DVFS: 700 MHz, 1.01 V DVFS: 920 MHz, 1.11 V DVFS: 1.2 GHz, 1.27 V 16 < 3% run to run variation across 10 runs; use the median Built a current sensing circuitry to measure the voltage and energy of the SoC (isolate from other board peripherals)
Hardware Setup Big core: ARM Cortex A9: OoO with 4 issue (e.g. in Tegra 3-based tablets) PandaBoard ES Rev B1, 45nm DVFS: 350 MHz, 0.83 V DVFS: 700 MHz, 1.01 V DVFS: 920 MHz, 1.11 V DVFS: 1.2 GHz, 1.27 V 17 Little core: ARM Cortex A8: In-order with 2 issue (e.g. in Apple A4 -- iPhone 4) BeagleBoard xM, 45 nm DVFS: 300 MHz, 0.94 V DVFS: 600 MHz, 1.10 V DVFS: 800 MHz, 1.26 V
21 “Webpages variance” in load time and energy Why Big/Little Systems? Different operating frequencies Different uarchitectures www.adobe.com www.newegg.com www.autoblog.com
Our Approach Workload Characterization Performance/Energy Prediction Resource Management How to leverage the big/little system to capture the webpage variance? 22
Webpage Characterization We treat webpages as the workload instead of the browser! HTML (Structure) CSS (Style) Selector Property Tag (h3, li, table, img) Attribute 24
Webpage Characterization DOM Tree We treat webpages as the workload instead of the browser! HTML (Structure) CSS (Style) Selector Property Tag (h3, li, table, img) Attribute 24
Webpage Characterization DOM Tree We treat webpages as the workload instead of the browser! HTML (Structure) CSS (Style) Selector Property Tag (h3, li, table, img) Attribute Tag (h3, li, table, img) HTML (Structure) 24
HTML Tag Analysis # Tags Webpages have a few hot HTML tags (hot instructions) # Tags 30 5K sorted by #tags Webpages have different tag counts (instruction counts) 5 Number of Tags (K)
HTML Tag Analysis # Tags Webpages have a few hot HTML tags (hot instructions) # Tags 30 5K sorted by #tags Webpages have different tag counts (instruction counts) 5 Number of Tags (K)
HTML Tag Analysis # Tags # Tags 31 5K sorted by #tags Webpages have a few hot HTML tags (hot instructions) Webpages have different tag counts (instruction counts) 5 Number of Tags (K)
HTML Tag Analysis # Tags # Tags 31 5K sorted by #tags Webpages have a few hot HTML tags (hot instructions) Webpages have different tag counts (instruction counts) 5 Number of Tags (K)
Tag Processing Overhead 32 0 50 100 150 200 h3 table img 0 175 350 525 700 Load time Energy ms mJ Webpages have a few hot HTML tags (hot instructions) Webpages have different tag counts (instruction counts)
Tag Processing Overhead 33 0 50 100 150 200 h3 table img 0 175 350 525 700 Load time Energy ms mJ Webpages have a few hot HTML tags (hot instructions) Webpages have different tag counts (instruction counts)
34 0 50 100 150 200 h3 table img 0 175 350 525 700 Load time Energy ms mJ Tag Processing Overhead Webpages have a few hot HTML tags (hot instructions) Webpages have different tag counts (instruction counts)
35 0 50 100 150 200 h3 table img 0 175 350 525 700 Load time Energy ms mJ Tag Processing Overhead Webpages have a few hot HTML tags (hot instructions) Webpages have different tag counts (instruction counts)
35 0 50 100 150 200 h3 table img 0 175 350 525 700 Load time Energy ms mJ HTML tags have different processing overhead (time & energy) Tag Processing Overhead Webpages have a few hot HTML tags (hot instructions) Webpages have different tag counts (instruction counts)
Characterization Conclusions 1. Webpages have different HTML tag counts and mixes 2. Individual HTML tags involve different processing overheads • Root cause of load time and energy variance 36
Regression Modeling Strategy Identify Training Predictors and Responses Training using hottest 2,500 webpages Model Construction and Refinement Start from the linear model and progressively refine it Model Validation Validating on another 2,500 webpages 39
Regression Modeling 40 Group Model Predictors Number of each tag HTML Number of each attribute Number of DOM tree nodes Number of rules CSS Number of each selector pattern Number of each property Content- dependent Total image size Content- dependent Total webpage size
Regression Modeling 40 Group Model Predictors Number of each tag HTML Number of each attribute Number of DOM tree nodes Number of rules CSS Number of each selector pattern Number of each property Content- dependent Total image size Content- dependent Total webpage size
Regression Modeling Group Model Predictors Number of each tag HTML Number of each attribute Number of DOM tree nodes Number of rules CSS Number of each selector pattern Number of each property Content- dependent Total image size Content- dependent Total webpage size 41
Webpage-aware Scheduling 48 Prediction (Minimal overhead) ........ Predict the load time and energy for each conf. Normal webpage rendering Webpage-aware scheduling
59 0 10 20 30 40 OS (Big) OS (Little) WS OS 0 25 50 75 100 Cut-off Violations (%) Energy Savings (%) Perf-oriented strategy as the baseline Results (Big+Little)
60 0 10 20 30 40 OS (Big) OS (Little) WS OS 0 25 50 75 100 Cut-off Violations (%) Energy Savings (%) Perf-oriented strategy as the baseline Results (Big+Little)
Conclusions Webpage-inherent characterization • Webpages are drastically different in load time and energy consumption Platform-dependent load time/energy prediction • 94.3% and 93.6% accuracy, respectively Big/little scheduling to effectively utilize the hardware resources • Significant energy saving over the performance-oriented strategy • Improve energy and performance over the Big/Little OS DVFS strategy
High-Performance and Energy- Efficient Mobile Web Browsing on Big/Little Systems Yuhao Zhu (गᜏᚠ) Vijay Janapa Reddi Trinity Research Group The University of Texas at Austin
queeten
onurcanalp
espacio
gihoz_support
juliawrobel
hide2kano
matleenalaakso
signer
mukaken
oripsolob
takunology
lara
jcasabona
notwaldorf
garrettdimon
bermonpainter
rmw
philhawksworth
kneath
jlugia
csswizardry
zakiyama