trace-driven analysis of power proportionality in storage systems

29
Trace-Driven Analysis of Power Proportionality in Storage Systems Sara Alspaugh and Arka Bhattacharya

Upload: hedya

Post on 24-Feb-2016

35 views

Category:

Documents


0 download

DESCRIPTION

Trace-Driven Analysis of Power Proportionality in Storage Systems. Sara Alspaugh and Arka Bhattacharya. Why trace-driven analysis. Lots of published proposals Giant design space. Some r elated work. Method. Laboratory. Production. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Trace-Driven Analysis of Power Proportionality in Storage Systems

Trace-Driven Analysis of Power Proportionality in Storage SystemsSara Alspaugh and Arka Bhattacharya

Page 2: Trace-Driven Analysis of Power Proportionality in Storage Systems

Why trace-driven analysis

• Lots of published proposals

• Giant design space

Page 3: Trace-Driven Analysis of Power Proportionality in Storage Systems

Some related workScheme Block

Device / RAID Level

File System Level

Fixed Thresh-hold

Pred-ictive

Erasure Codes (RAID5)

Mirror-ing (RAID1)

Write Logging

Access Freq.-Based Layout

Solid State Devices

Multi-speed Disks

Hybrid / Tiered

DIV-ACC X X X X

EERAID1 X X X X

EERAID5 X X X X X

RIMAC X X X X

PARAID X X X X

PDC X X X

PA-LRU X X X

PB-LRU X X X X

HIBERN X X X X X X

DPRM X X X X

WOL X X X X X X

MAID X X X X

SSD-RAID X X X X X X

EED X X X X X X

SIERRA X X X X

RABBIT X X X X X

Page 4: Trace-Driven Analysis of Power Proportionality in Storage Systems

Method

EvaluationLaboratory Production Implementation is

infeasible when considering many system types.

AnalysisComponents

Traces

Algorithms

?

Page 5: Trace-Driven Analysis of Power Proportionality in Storage Systems

Trace Type Citation

Wikipedia HTTP SOCC ‘10NetApp, Harvard NFS USENIX ‘08, LISA ‘03MSR Cambridge Block Device FAST ‘08Facebook Analytics Hadoop MapReduce EuroSys ‘11Google Web Search ISCA ‘11

Page 6: Trace-Driven Analysis of Power Proportionality in Storage Systems

AnalysisComponents

Traces

Algorithms

CharacteristicsRequest RateInterarrival TimesRead-Write Mix...

Page 7: Trace-Driven Analysis of Power Proportionality in Storage Systems

Quantifying Inherent Opportunity• gain =

diff(peak x length, sum(bandwidth)) /peak x length

• waste factor = peak x length / sum(bandwidth)

• waste factor = peak:avg

Page 8: Trace-Driven Analysis of Power Proportionality in Storage Systems

time

band

widt

h

Page 9: Trace-Driven Analysis of Power Proportionality in Storage Systems
Page 10: Trace-Driven Analysis of Power Proportionality in Storage Systems
Page 11: Trace-Driven Analysis of Power Proportionality in Storage Systems
Page 12: Trace-Driven Analysis of Power Proportionality in Storage Systems

time

band

widt

h

Page 13: Trace-Driven Analysis of Power Proportionality in Storage Systems

data set size (B)

band

widt

h re

quire

men

ts

(B/s

)

data set size (B)

band

widt

h re

quire

men

ts (B

/s)

data set size (B)ba

ndwi

dth

requ

irem

ents

(B

/s)

data set size (B)

band

widt

h re

quire

men

ts

(B/s

)

bw_app >> bw_componentcap_app < cap_component

bw_app <= bw_{components}cap_app >> cap_component

Page 14: Trace-Driven Analysis of Power Proportionality in Storage Systems

unit = disks

Band

widt

h (b

ytes

/ se

c )

Capacity (bytes)

partition

replicate

~ 500 GB

~ 50 MB/s

laptop NFS filer

DB server

Page 15: Trace-Driven Analysis of Power Proportionality in Storage Systems

unit = servers

band

widt

h

bytes

partition

replicate

~ 12 TB (disk)

memory cache

DFS

~ 32 GB (RAM)

DB server

~ 200 MB/s

~ 1 GB/s

Page 16: Trace-Driven Analysis of Power Proportionality in Storage Systems
Page 17: Trace-Driven Analysis of Power Proportionality in Storage Systems
Page 18: Trace-Driven Analysis of Power Proportionality in Storage Systems

data set size (B)ba

ndwi

dth

requ

irem

ents

(B/s

)

data set size (B)

band

widt

h re

quire

men

ts

(B/s

)

data set size (B)

band

widt

h re

quire

men

ts

(B/s

)

NAS / NFS (NetApp), disk arrays

web farms (Wikipedia)

data analytics, DFS (Hadoop)

Page 19: Trace-Driven Analysis of Power Proportionality in Storage Systems

data set size (B)

band

widt

h re

quire

men

ts

(B/s

)

data set size (B)

band

widt

h re

quire

men

ts (B

/s)

data set size (B)ba

ndwi

dth

requ

irem

ents

(B

/s)

data set size (B)

band

widt

h re

quire

men

ts

(B/s

)

bw_app >> bw_componentcap_app < cap_component

bw_app <= bw_{components}cap_app >> cap_component

Page 20: Trace-Driven Analysis of Power Proportionality in Storage Systems

Challenges• Case 1: writes• Case 2: latency to inactive

components• Case 3: both of the above, set cover

problem

Page 21: Trace-Driven Analysis of Power Proportionality in Storage Systems
Page 22: Trace-Driven Analysis of Power Proportionality in Storage Systems
Page 23: Trace-Driven Analysis of Power Proportionality in Storage Systems

write through: to all components (even if requires waking some)

write offloading: to active components only (propagate on wake)

write log: propagate when ~full reaper: to all components but only wake when queue is full

Page 24: Trace-Driven Analysis of Power Proportionality in Storage Systems

time

band

widt

h

requests

active units write-offloading

active units write-through

Page 25: Trace-Driven Analysis of Power Proportionality in Storage Systems
Page 26: Trace-Driven Analysis of Power Proportionality in Storage Systems
Page 27: Trace-Driven Analysis of Power Proportionality in Storage Systems
Page 28: Trace-Driven Analysis of Power Proportionality in Storage Systems

Next steps• data not pictured

here– latencies– ramp times– unit sizes– etc.

• ways to slice it• how to visualize it

• more workloads• go back to related

work to compare• case 3– object popularity

Page 29: Trace-Driven Analysis of Power Proportionality in Storage Systems

QUESTIONS?The End.