SLIDE 1
CAMELab ab
Motivation
2
Host
..
Garbage Collection
SSD
NAND Flash density (pages / block)
More GC overhead
Delay
How to hide GC latency?
- Let’s perform GCs at user idle times!
CA CAME MELab ab Motivation Host NAND Flash density (pages / - - PowerPoint PPT Presentation
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
SLIDE 2
SLIDE 3
CAMELab ab
Hiding GC latency : Background GC
3
SSD
I/O Garbage Collection Time Wait
Common Assumption:
Storage won’t be touched in the near future! Request Threshold 1 10 100 1k 10us 10ms 10s Idle Time
# Request
SLIDE 4
CAMELab ab I/O
Hiding GC latency : Background GC
*Real workload from MS Production Server (https://trace.camelab.org/) 4
SSD
Garbage Collection Time Wait Request
1 100 10k 10s 10ms 10us
Assumption Real workload*
I/O Delay! 1 10 100 1k 10us 10ms 10s
# Request
Threshold
SLIDE 5
CAMELab ab
GC-Tutor
5
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
SLIDE 6
CAMELab ab
DNN-based GC Scheduling
6
I/O Pattern DNN-based Idle Time Prediction Background GC
Problem :
A fixed DNN model fails to predict unseen workloads
DNN Model
Timestamp, R/W, seq/rand, size
Idle time
SLIDE 7
CAMELab ab
DNN-based GC Scheduling
7
Problem :
A fixed DNN model fails to predict unseen workloads
Online Learning!
I/O Pattern DNN-based Idle Time Prediction Background GC
1D-CNN Model
Timestamp, R/W, seq/rand, size
Idle time
SLIDE 8
CAMELab ab
Lightweight Online Learning
*Chelsea Finn, et al., Model Agnostic Meta Learning for Fast Adaptation of Deep Networks, ICML 2017 8
25 50 75 100
Accuracy (%) Deeplearn GCTutor
Offline Online
Meta Learning I/O traces
Online Learning wdev prxy stg
Meta Learning* Naïve
Takes more than a few hours
Infeasible!
SLIDE 9
CAMELab ab
Evaluation
Traces from CAMELab Trace(https://trace.camelab.org/) 9
C F S 2 4 H R B S D D R D A P
- n
l i n e w e b m a i l r s r c h 10 60 70 80 90 100
Accuracy (%)
Train set prn proj prxy stg wdev 20 40 60 80 100
Accuracy False short False long
Unseen set
Left: GCTutor Right: Deep
GC-Tutor can accurately predict idle time
- Consistently higher accuracy on trained workloads
- Significantly higher accuracy on unseen workloads
- prxy, stg :
Very different idle time distribution compared to trained workloads
GC-Tutor can reduce GC-induced delays by 82.4%,
- n average, compared to rule-based GC scheduler
SLIDE 10
CAMELab ab
Conclusion : GC-Tutor
10
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!
Offline Online
Meta Learning
I/O traces
Online Learning
25 50 75 100
Accuracy (%) GCTutor Deeplearn
wdev prxy stg
SLIDE 11
CAMELab ab
Thank you!
Junhyeok Jang Electrical Engineering, KAIST
11