18
Reducing Smartphone Application Delay through Read/Write Isolation David T. Nguyen

Reducing Smartphone Application Delay through Read/Write Isolation

Embed Size (px)

Citation preview

Storage-aware Smartphone Energy Savings

Reducing Smartphone Application Delaythrough Read/Write Isolation

David T. Nguyen

David T. Nguyen2Figure Courtesy: Facebook Connectivity Lab

Real-world Problem: Application Delay

David T. NguyenWilliam & Mary2

IntroductionResearch questionsHow does disk I/O affect smartphone application response time?How to optimize storage to improve application performance?

ContributionsFirst large-scale Android I/O study on 2611 devicesProposed SmartIO scheduler that reduces application delayEvaluation indicates 21% reduced launch and 17% reduced run-time delay

David T. Nguyen3

www.StoreBench.comStoreBench Benchmark4

David T. Nguyen

40% devices between 13% and 58% Significant CPU time spent waiting for storage

Iowait of 2611 DevicesCPU active time spent waiting for storage I/O5David T. Nguyen

David T. NguyenWilliam & Mary5

Seq. reads experience up to 626% slowdown when blocked by concurrent writes (rand. up to 293%)I/O slowdown possible source of worse app launch (reads dominate)

I/O Slowdown6

David T. Nguyen

Slowdown Asymmetry 7

ReadSlowdown = Response time of a read in the presence of a concurrent write / Response time of a read when running aloneWriteSlowdown = Response time of a write in the presence of a concurrent read / Response time of a write when running aloneRead slowdown up to 6.15 vs. write slowdown up to 1.6

David T. Nguyen

Study SummaryAndroid devices spend up to 58% of CPU active life time idling due to I/O

Reads experience up to 626% slowdown when blocked by concurrent writes

Large slowdown asymmetry (reads up to 6.15; writes up to 1.6)8David T. Nguyen

I/O Scheduler Storage SubsystemConcurrency ProfilerI/O GroupingI/O Dispatch

I/O Priority Assignment

Kernel SpaceUser Space

SmartIODavid T. Nguyen9

9

FLASH BLOCK DEVICE

0 1 2 3 4 5 6 7

0 1 2 3 4 5 6 7 7

queuequeue

IdleBest EffortReal-time

writereadDispatchQueue Real-timeBest EffortIdle

10David T. Nguyen

11David T. Nguyen

Demo SmartIO11

Nexus 4: 1-reader (128MB), 1-writer (128MB)Seq. read slowdown improves from 6.15 to1.72 (gain), write worsens from 1.13 to 1.51 (cost)Rand. read slowdown improves from 3.18 to 1.97 (gain), write worsens from 1.6 to 1.83 (cost)

Gain vs. Cost12

David T. Nguyen

128MB mixed workload (first 10% reads, then 20%..)With increasing percentage of reads, response time decreases (less writes means less slowdown)By changing scheduler (default CFQ to SmartIO) we achieve on average 49% faster response (max. 66%)

Scheduler Comparison13

David T. Nguyen

Cold launch delays reduced on average by 21% (warm by 13%)Run-time delays reduced by 17%

14

David T. Nguyen

Facebook Inc.s goal: cold launch < 5s on 2012 or newer devicescold launch < 10s on older devicesRAZR (2012): 9.9s -> 6.2sSamsung S5 (2014): 3.7s -> 2.3sFacebook Performance15

David T. Nguyen

Conclusion & Future WorkSmartIOReduces app delays through read/write isolation and higher read priorities

Study other stages of app life cycle (install, update, switch, uninstall)

Analyze impact of network I/O16David T. Nguyen

AcknowledgmentsOur shepherd Prof. Mahadev Satyanarayanan (CMU)Aaron Carroll (NICTA), Dr. Duy Le (EMC), Dr. Tommy Nguyen (RPI), Mai Anh Do (CNU), Daniel Graham (W&M)17

David T. Nguyen

Reducing Smartphone Application Delaythrough Read/Write Isolation

David T. Nguyen