Online Phase-Adaptive Data Layout Selection Chengliang Zhang - - PowerPoint PPT Presentation

▶
online phase adaptive data layout selection
SMART_READER_LITE
LIVE PREVIEW

Online Phase-Adaptive Data Layout Selection Chengliang Zhang - - PowerPoint PPT Presentation

Nov 08, 2023 169 likes •346 views

Online Phase-Adaptive Data Layout Selection Chengliang Zhang Martin Hirzel Microsoft IBM (former IBM intern) ECOOP, 10 July 2008 1 Problem Statement No training run. Online Phase-Adaptive Data Layout Selection Cache line or page A

slide-1
SLIDE 1

1

Online Phase-Adaptive Data Layout Selection

ECOOP, 10 July 2008

Chengliang Zhang Martin Hirzel

Microsoft (former IBM intern)

IBM

slide-2
SLIDE 2

2

Online Phase-Adaptive Data Layout Selection

Cache line or page A Cache line or page B Object 1 Object 2 Object 3 Object 4

try measure decide

No training run.

Problem Statement

slide-3
SLIDE 3

3

Data Layouts from Copying Garbage Collection

BF HI

slide-4
SLIDE 4

4

Layout Performance Comparison

8-processor AMD HI faster BF faster

slide-5
SLIDE 5

5

Multi-Armed Bandit Problem

BF HI Depth-first Allocation order Popularity Random Size Type Thread

slide-6
SLIDE 6

6

Layout Auditing

try Data reorganizer Program measure Profiler decide Controller

Performance Data Layout Perfor- mance Profiling Decision Layout Decision Reward

slide-7
SLIDE 7

7

Profiler

measure Profiler

Profiling Decision Reward

Data reorg. Program Data reorg. Program Data reorg. Program Data layout

l3 l4 l5

Physical time (wall clock)

r3 e3 r4 e4 r5 e5

Virtual time (allocated bytes)

v3 v4 v5

Reward for layout li uses historical average of:

  • Virtual time vi / program execution time ei
  • Virtual time vi-1 / reorganizer time ri

(always on)

slide-8
SLIDE 8

8

Controller: Blind Justice

Profiling Decision Layout Decision Rewards

Goals

  • Match performance
  • f best layout
  • Online

Challenges

  • Confidence
  • vs. Curiosity
  • Phase changes
  • vs. Noise
slide-9
SLIDE 9

9

Confidence vs. Curiosity

Confidence Curiosity Never tried layout Few samples / High variance Many samples / Low variance

∞

Pick layout l if either:

  • High confidence that l gives best reward
  • High curiosity about l ’s reward

⇒ use simulated annealing

slide-10
SLIDE 10

Phase Changes vs. Noise

Phase Adaptivity

– When layout performance changes, learn new best layout ⇒Forget historical rewards

Noise Tolerance

– Perturbation from extraneous causes ⇒ Remember historical rewards

10

⇒ use exponential decay

slide-11
SLIDE 11

11

SASO Properties of Control Systems

  • Stability
  • Accuracy
  • Settling
  • Overshoot
  • Phase adaptivity
  • Overhead
slide-12
SLIDE 12

Methodology

12

20 Java programs

(DaCapo suite, SPECjvm98 suite, and a few more)

J9 = IBM’s product Java VM

HI

hierarchical

BF

breadth-first

LA

layout auditing

4 Hardware Platforms Intel-2 AMD-2 AMD-4 AMD-8

slide-13
SLIDE 13

Accuracy and Overhead

13

1 2 Intel

  • 2

AMD-2 AMD-4 AMD-8 BF HI LA

Average % slowdown

  • vs. best

5 10 Intel

  • 2

AMD-2 AMD-4 AMD-8 BF HI LA 10 20 Intel

  • 2

AMD-2 AMD-4 AMD-8 BF HI LA

Number of programs not optimal Worst % slowdown

  • vs. best
slide-14
SLIDE 14

Stability and Settling

14

Decay = 0.9 No decay (Decay=1.0)

75s 20s

BF better HI better HI better BF better HI better HI better Reward Layout Decision

BF HI

slide-15
SLIDE 15

15

Related Work

  • Lau/Arnold/Hind/Calder PLDI’06:

performance auditing for JIT optimization

  • Soman/Krintz/Bacon ISMM’04: switch

copy vs. mark-sweep, generations or not

  • Chen/Bhansali/Chilimbi/Gao/Chuang

PLDI’06: throttle unless miss rate reduced

  • Saavedra/Park PACT’96: adapt prefetch

distance based on cancellation & latency

slide-16
SLIDE 16

16

  • Accurate
  • Phase adaptive (good settling/stability)
  • Negligible overhead profiling
  • Online, hardware independent

Conclusions

slide-17
SLIDE 17

17

Clustering Layouts by Performance

[SIGMETRICS 2007]

BF HI