CA CAME MELab ab Motivation Host NAND Flash density (pages / - PowerPoint PPT Presentation

Platform-Agnostic Lightweight Deep Learning for Garbage Collection Scheduling in SSDs Junhyeok Jang, Donghyun Gouk, Jinwoo Shin, Myoungsoo Jung Computer Architecture and Memory systems Laboratory CA CAME MELab ab

Motivation Host NAND Flash density (pages / block) Delay More GC overhead .. How to hide GC latency? • Let’s perform GCs at user idle times! How long will be Garbage Collection the user idle times? SSD CAMELab ab 2

Hiding GC latency : Background GC Request Time I/O Garbage Collection SSD Wait Threshold Idle Time 1k # Request Common Assumption: 100 10 Storage won’t be touched 1 10us 10ms 10s in the near future! CAMELab ab 3

Hiding GC latency : Background GC Request Time I/O Garbage Collection I/O SSD Delay! Wait Threshold 1k 10k # Request 100 100 10 1 1 10us 10ms 10s 10us 10ms 10s Assumption Real workload* CAMELab ab 4 *Real workload from MS Production Server (https://trace.camelab.org/)

GC-Tutor DNN-based GC scheduler • Precisely predict future request arrivals • Schedules GC in user-invisible time • Consistently accurate regardless of workload with lightweight online learning mechanism CAMELab ab 5

DNN-based GC Scheduling DNN Model Idle time I/O Pattern Timestamp, R/W, seq/rand, size DNN-based Idle Time Prediction Background GC Problem : A fixed DNN model fails to predict unseen workloads CAMELab ab 6

DNN-based GC Scheduling 1D-CNN Model Idle time I/O Pattern Timestamp, R/W, seq/rand, size DNN-based Idle Time Prediction Background GC Problem : A fixed DNN model fails to predict unseen workloads Online Learning! CAMELab ab 7

Lightweight Online Learning Naïve Meta Learning* Offline Online wdev prxy stg 100 Accuracy (%) I/O traces 75 50 Meta 25 Deeplearn Learning GCTutor Takes more than a few hours 0 Infeasible! Online Learning CAMELab ab 8 *Chelsea Finn, et al., Model Agnostic Meta Learning for Fast Adaptation of Deep Networks, ICML 2017

Evaluation Train set GC-Tutor can accurately predict idle time Accuracy (%) 100 90 • Consistently higher accuracy on trained workloads 80 • Significantly higher accuracy on unseen workloads 70 60 • prxy, stg : 10 0 Very different idle time distribution compared to trained workloads S R S R P e l h i a n c F B A H D m r i C D l 4 D s n b r 2 o e w Unseen set 100 80 False long False short Accuracy 60 40 20 0 GC-Tutor can reduce GC-induced delays by 82.4%, prn proj prxy stg wdev on average, compared to rule-based GC scheduler Left: GCTutor Right: Deep CAMELab ab 9 Traces from CAMELab Trace(https://trace.camelab.org/)

Conclusion : GC-Tutor Offline Online wdev prxy stg 100 Accuracy (%) I/O traces 75 50 Meta GCTutor 25 Learning Deeplearn 0 Online Learning DNN-based GC scheduler • Accurate request arrival prediction using DNN model • Meta learning-based light-weight online learning mechanism Making GC overhead invisible to users! CAMELab ab 10

Thank you! Junhyeok Jang Electrical Engineering, KAIST CAMELab ab 11