exploiting managed language semantics to mitigate wear-out ... · 07/08/2019 · gc to limit pcm...

Exploiting Managed Language Semantics to Mitigate Wear-out in Persistent Memory

Shoaib AkramGhent University, Belgium

Flash Memory Summit Santa Clara, CA

20191

Main memory capacity expansion


20192

Charge storage in DRAM a scaling limitation

Manufacturing complexity makes DRAM pricing volatile

Source: WSTS, IC Insights0.6

0.7

0.8

0.9

1

Pric

e/Gb

($)

Jan’17 Jan’18

Phase change memory (PCM)


20193

Scalable → More Gb for the same price

Byte addressable like DRAM

Latency closer to DRAM

🙁🙁 Low write endurance

Why PCM has low write endurance?


20194

tem

pera

ture

set to crystalline

reset to amorphous

time

Electric pulses to program PCMcells wear them out over time

Hybrid DRAM-PCM memory


20195

DRAM PCM

Endurance

PCM alone as a DRAM replacement wears out in a few months for popular Java applications

CapacityPersistence

Mitigating PCM wear-out


20196

Wear-leveling to spread writes across PCM

This talk → Use DRAM to limit PCM writes

OS to limit PCM writes


20197

Coarse-grained page migrations hurt application performance and PCM lifetime

DRAM PCM


20198

Managed runtimes

Managed Runtime

OperatingSystem

Hardware

ApplicationPlatform independenceAbstract hardware/OS→ Aka Virtual Machine

Ease programmer’s burdenGarbage collectionSecurity


20199

GC to limit PCM writes

GC understands memory semantics

OperatingSystem

Hardware

Application

GC approaches are pro-activeand fine-grained


201910

Write Distribution in GC heap

nursery matureGC

70%of writes


201911

Write Distribution in GC heap

22%to 2% of objects

nursery matureGC

70%of writes


201912

Write-Rationing Garbage Collection

Limit PCM writes by discovering highly written objects

Kingsguard dynamically monitor writes

Kingsguard-Nursery (KG-N)


201913

mature largenursery

DRAM PCM

Kingsguard-Writers (KG-W)


201914

mature large

observer

PCM

mature largeDRAM

nursery

KG-W drawbacks


201915

Overhead of dynamic monitoring

Limited time window to predict write intensity

Excessive & fixed DRAM consumption


201916

Write-Rationing Garbage Collection

Limit PCM writes by discovering highly written objects

Kingsguard dynamically monitor writes

Crystal Gazer statically profiles objects

Allocation site as a write predictor


201917

a = new Object()b = new Object()c = new Object()d = new Object()

a = new Object()b = new Object()c = new Object()d = new_dram Object()

Uniform distribution 🙁🙁Skewed distribution 🙂🙂

Write distribution by allocation site


201918

Few sites capture majority of writes

0

25

50

75

100

0 50 100 150

% m

atur

e ob

ject

s

Sites sorted by writes

WritesVolume Pjbb2005

Crystal Gazer operation


201919

ApplicationProfiling

Advice Generation

Bytecode Compilation

a = new Object()…b = new_dram Object()

a = new Object()…b = new Object()

ObjectPlacement

Advice generation


201920

Goal: Generate <alloc-site, advice> pairs advice → DRAM or PCMinput is a write-intensity trace

Two heuristics to classify allocation sites as DRAM or PCM

Classification heuristic (1)


201921

Freq: A threshold % of objects from a site get more than a threshold # writes → DRAM

🙂🙂 Aggressively limits PCM writes

🙁🙁 No distinction based on object size

Classification heuristic (2)


201922

Dens: A threshold % of objects from a site have more than a threshold write density → DRAM

Write density → Ratio of # writes to object size

Classification thresholds


201923

Homogeneity threshold → 1%

Frequency threshold → 1

Density threshold → 1

Classification examples


201924

ObjectIdentifier # Writes # Bytes

Allocation site

O1 0 4 A() + 10O2 0 4 A() + 10O3 128 4 A() + 10O4 128 4096 B() + 4

Frequency threshold = 1PCM writes = ?, DRAM bytes = ?



201925


Allocation site

O1 0 4 A() + 10O2 0 4 A() + 10O3 128 4 A() + 10O4 128 4096 B() + 4

Frequency threshold = 1PCM writes = ?, DRAM bytes = ?

→ →



201926


Allocation site

O1 0 4 A() + 10O2 0 4 A() + 10O3 128 4 A() + 10O4 128 4096 B() + 4

Frequency threshold = 1PCM writes = 0/256, DRAM bytes = 5008

→ →



201927


Allocation site

O1 0 4 A() + 10O2 0 4 A() + 10O3 128 4 A() + 10O4 128 4096 B() + 4

Density threshold = 1PCM writes = ?, DRAM bytes = ?



201928


Allocation site

O1 0 4 A() + 10O2 0 4 A() + 10O3 128 4 A() + 10O4 128 4096 B() + 4


→ 32



201929


Allocation site

O1 0 4 A() + 10O2 0 4 A() + 10O3 128 4 A() + 10O4 128 4096 B() + 4


→ <1



201930


Allocation site

O1 0 4 A() + 10O2 0 4 A() + 10O3 128 4 A() + 10O4 128 4096 B() + 4

Density threshold = 1PCM writes = 128/256, DRAM bytes = 12


201931

new_dram() → Set a bit in the object header

GC → Inspect the bit on nursery collection to copy object in DRAM or PCM

Object placement in Crystal Gazer


201932

Object placement in Crystal Gazer

mature large

PCM

mature largeDRAM

nursery🧐🧐

Is marked highly written? ✓

Key features of Crystal Gazer


201933

Eliminates overhead of dynamic monitoring

Less mispredictions due to pro-active nature

Pareto optimal trade-offs b/w capacity and lifetime


201934

Evaluation methodology

15 Applications → DaCapo, GraphChi, SpecJBB

Medium-end server platform

Different inputs for production and advice

Jikes RVM


201935

Emulation platform

CPU CPU✗

Jikes RVMApp

OS


201936

PCM write rates

PCM-Only write rate is above 1 GB/s on average

Safe operation is 200 MB/s


201937

PCM write rates

0200400600800

Writ

e ra

te in

MB/

s KG-N KG-W Dens Freq


201938

Execution time

0.0

0.5

1.0

1.5Ex

ec. t

ime

norm

. to

KG-N

KG-W Dens Freq

30% 8%


201939

0.30.40.50.60.70.8

100 150 200 250

PCM

writ

es

rela

tive

to K

G-N

DRAM capacity in MB

Pjbb2005

KG-W

KG-W versus Crystal Gazer


201940

0.30.40.50.60.70.8

100 150 200 250

PCM

writ

es

rela

tive

to K

G-N

DRAM capacity in MB

Pjbb2005

KG-WCrystal Gazer

KG-W versus Crystal Gazer

Crystal Gazeropens up Pareto-optimaltrade-offs


201941

Write-rationing garbage collection

Hybrid memory is inevitable

All layers can contribute to manage hybrid memory

Write-rationing GC is pro-active and fine-grained

DRAM PCM

exploiting managed language semantics to mitigate wear-out ... · 07/08/2019 · gc to limit pcm...

Documents