Architecture Hugo M IOMANDRE , Julien H ASCOT , Karol D ESNOS , Kevin - - PowerPoint PPT Presentation

▶

Dec 27, 2022 260 likes •460 views

INSTITUT DLECTRONIQUE ET DE TLCOMMUNICATIONS DE RENNES Porting the Spider Dataflow Runtime on the Kalray MPPA Manycore Architecture Hugo M IOMANDRE , Julien H ASCOT , Karol D ESNOS , Kevin M ARTIN , Benoit D UPONT DE D INECHIN

SLIDE 1

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

Hugo MIOMANDRE, Julien HASCOËT, Karol DESNOS, Kevin MARTIN, Benoit DUPONT DE DINECHIN Jean-François NEZAN

Porting the Spider Dataflow Runtime

n the Kalray MPPA Manycore

Architecture

Dataflow Workshop - 2017.12.12

SLIDE 2

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

Reconfigurable Dataflow for Manycore

Context > Overview

GdR ISIS Project

Archi model

Modeling Framework Runtime Adaptation Layer Manycore Architecture

SLIDE 3

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

The PiSDF Model of Computation

Context > Dataflow input

N

Read Header

Size

Size Size

Read Image Filter

/N Size

Size

in Size

SetNb Slices

Size

Kernel

/N Size Size Size Size

Send

Desnos et al. "Pimm: Parameterized and interfaced dataflow meta-model for mpsocs runtime reconfiguration." SAMOS, 2013

SLIDE 4

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

The SPIDER runtime manager

Context > Runtime

Heulot et al. "Spider: A synchronous parameterized and interfaced dataflow-based RTOS for multicore DSPs" EDERC, 2014 Jobs Jobs Jobs Params Timings Data Data

Pool of data FIFOs

Master tasks:

1. Manage graphs
2. Map & Schedule
3. Send Jobs
4. Run jobs
5. Monitor & Trace

Slave task:

Run jobs

Slave Slave Master

SLIDE 5

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

Manycore Architecture: Kalray’s MPPA256B

Context > The Beast

Challenges

Distributed scratchpad memory
Massive parallelism
NoC-Based communications

SLIDE 6

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

Distributed synchronization
Lightweight manycore scheduling
Dataflow-based distributed memory allocation

Contributions

Miomandre et al. "Embedded Runtime for Reconfigurable Dataflow Graphs on Manycore Architectures" submitted to PARMA-DITAM 18

SLIDE 7

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

Contributions > Overview

Spider distribution on MPPA

Master Slave X 256

SLIDE 8

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

Contributions > Synchro

Previous Synchronization Mechanisms

Shared-Memory Based

(x86, Big.LITTLE)

Core1 Flag Core2

A B Poll

Pop job

Hardware supported

(TI Keystone 2)

Core1 FIFO Core2

A B

Pop job

SLIDE 9

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

Contributions > Synchro

Synchronization Mechanisms

Before: Centralized Polling Master Slave Slave Slave NoC After: Distributed notifications NoC Master Slave Slave Slave

Unbounded

communications

Contention
Notifier/Observer pattern
Bounded

communications

SLIDE 10

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

Contributions > Scheduling

Complexity issue for manycore

Good’ol Lis ist t sche heduli duling ng

1. Create a sorted list of

all actors to map/schedule.

2. Schedule first actor
f the list of first

available core.

3. Go back to step 2

until the list is empty.

Complexity: O( A.log(A) + P.A))

A: # actors P: # processors Multicore architectures: P smaller than log(A) => A.log(A) dominates Manycore architecture: P is large A ∝ P (by designer) => Complexity quadraticaly

SLIDE 11

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

Contributions > Scheduling

Specialized Round Robin

1.Create a list of actor in topological order. 2.Schedule first actor by systematically interleaving clusters and cores 3.Go back to step 2 until the list is empty.

Lightweight Scheduling

Complexity: O( A )

A: # actors Advantages:

Low complexity
Interleaving clusters avoids

communication contentions between starting actors.

The scheduler controls the number
f jobs in the queue of a core.

SLIDE 12

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

Contributions > Memory Alloc.

Classic Memory Allocation

Malloc In & Out Success? Execute actor

OK KO

Crash!

SLIDE 13

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

Contributions > Memory Alloc.

Cluster-Level Dataflow Aware Memory Allocation

Malloc In & Out Success? Execute actor

Notif. Master

OK KO

Other Actor?

OK OK

SLIDE 14

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

Results > Application

+ Morphological operator

SLIDE 15

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

Results > Performance

Max speedup: 22 on 256 cores (on 4k video)

SLIDE 16

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

Results > Energy

2.5 more energy efficient (On 4k video)

Intel Xeon E5-1650 6-hyperthraded x86 Kalray MPPA256 Bostan 258 RISC Cores 11.40 fps 2.81 fps ~120W ~12W

SLIDE 17

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES

Summary

Successful porting of Spider on MPPA

Feasible: Reconfigurable Dataflow Runtime for manycore
Promising: Energy efficiency beats x86
Open source(-ish): On github!

Future Work

Better
Faster
Stronger

Julien, Florian, Alexandre, Hamza, Claudio, …

SLIDE 18

K. Desnos – SPIDER / MPPA

INSTITUT D’ÉLECTRONIQUE ET DE TÉLÉCOMMUNICATIONS DE RENNES