Sambamba : A Runtime System for Online Adaptive Parallelization - - PowerPoint PPT Presentation

▶

Feb 10, 2023 203 likes •383 views

Sambamba : A Runtime System for Online Adaptive Parallelization Clemens Hammacher Kevin Streit Sebastian Hack Andreas Zeller A well-known challenge Modern Legacy Code Hardware The Problem ... The Problem ... Thread 1 Thread 2 Thread 3

SLIDE 1

Sambamba: A Runtime System for Online Adaptive Parallelization

Clemens Hammacher Kevin Streit Sebastian Hack Andreas Zeller

SLIDE 2

A well-known challenge

Legacy Code Modern Hardware

SLIDE 3

The Problem

...

SLIDE 4

The Problem

...

Thread 1 Thread 2 Thread 3 Thread 4 Thread 5

SLIDE 5

The Problem

... ...

Thread 1 Thread 2 Thread 3 Thread 4 Thread 5

SLIDE 6

The Problem

... ...

Thread 1 Thread 2 Thread 3 Thread 4 Thread 5 Thread 1 Thread 2 Thread 3 Thread 4

SLIDE 7

The Problem

... ...

Thread 1 Thread 2 Thread 3 Thread 4 Thread 5 Thread 1 Thread 2 Thread 3 Thread 4

input dependent
machine dependent

SLIDE 8

Compile-time Analyses Parallelization Profiling Adaptation Execution

Compute data dependences (using DSA)
Generate program dependence graph
Precompute parallelization candidates

SLIDE 9

Compile-time Analyses Parallelization Profiling Adaptation Execution

Combine local parallel schedules

SLIDE 10

Compile-time Analyses Parallelization Profiling Adaptation Execution

JIT
compile code and inject into application
Decide which code variant to execute

SLIDE 11

Compile-time Analyses Parallelization Profiling Adaptation Execution

Profile execution times
Profile branch probabilities

SLIDE 12

Compile-time Analyses Parallelization Profiling Adaptation Execution

Analyze gathered profiles
Trigger reassessing parallelization decisions

SLIDE 13

Compile-time Analyses Parallelization Profiling Adaptation Execution

Analyze gathered profiles
Trigger reassessing parallelization decisions

SLIDE 14

Compile-time Analyses Parallelization Profiling Adaptation Execution

Analyze gathered profiles
Trigger reassessing parallelization decisions

SLIDE 15

Speedup on CILK example programs

0X 1X 2X 3X 4X 5X

All experiments run on an i7 quad-core with HT.

SLIDE 16

Compile-time Analyses Parallelization Profiling Adaptation Execution

Sambamba

Clemens Hammacher Kevin Streit Sebastian Hack Andreas Zeller