서울대학교

Software Platform Lab 소개

전병곤

2021년 10월 15일

§ AI 및 빅데이터시스템분야최고연구수행, 초거대 AI 모델연구수행

§ 리더전병곤교수와 1명박사후연구원, 8명박사과정, 9명석사과정으로구성

§ 2021년여름까지 5명의박사졸업생, 6명의석사졸업생배출

§ 초거대 AI 개발플랫폼스타트업 FriendliAI 창업

§ 마이크로소프트, 페이스북, 구글, 삼성, 네이버등글로벌기업과협업, 인턴십

§ 아파치재단등을포함활발한공개소프트웨어활동참여

Software Platform Lab은?

전병곤교수약력

!"# $%

• 서울대컴퓨터공학부교수

• FriendliAI 대표

• 서울대 AI�연구원 부원장

• 서울대-네이버초거대(Hyperscale) AI연구 센터장

• 초거대 AI플랫폼기술개발

• (전) MS,�Yahoo!,�Intel�

• (전) Facebook,�네이버 방문연구원

• UC�Berkeley 박사,�Stanford�석사, 서울대 석사, 학사

• 지난 10년간 가장 impact있는 시스템연구로유럽과미국에서수상 (2021,�2020)

• 다수의연구상수상:�구글(2020),�아마존(2018),�페이스북(2017),�마이크로소프트(2014)

• 논문 14,500회 이상인용됨 (Google�Scholar)

§ WindTunnel: Towards Differentiable ML Pipelines Beyond a Single Model (VLDB 2022)§ Terra: Imperative-Symbolic Co-Execution of Imperative Deep Learning Programs (NeurIPS2021)

§ Finding Consensus Bugs in Ethereum via Multi-transaction Differential Fuzzing (OSDI 2021)§ Refurbish Your Training Data: Reusing Partially Augmented Samples for Faster Deep Neural Network Training (ATC 2021)

§ Nimble: Lightweight and Parallel GPU Task Scheduling for Deep Learning (NeurIPS 2020 Spotlight)

§ Apache Nemo: A Framework for Building Distributed Dataflow Optimization Policies (ATC 2019)§ JANUS: Fast and Flexible Deep Learning via Symbolic Graph Execution of Imperative Programs (NSDI 2019)

§ Parallax: Sparsity-aware Data Parallel Training of Deep Neural Networks (EuroSys 2019)§ PRETZEL: Opening the Black Box of Machine Learning Prediction Serving Systems (OSDI 2018)§ Improving the Expressiveness of Deep Learning Frameworks with Recursion (EuroSys 2018)

최근 3년발표중요연구 (OSDI, NSDI, EuroSys, ATC, VLDB, NeurIPS 등)

https://spl.snu.ac.kr/publications/

§ (초거대) AI 시스템§ 컴파일러

§ 런타임

§ 수행모델

§ 분산처리

§ AutoML

§ AI 클러스터자원관리

§ 모바일 AI 시스템

AI Systems & Models, Data Processing Systems

§ 초거대 AI 모델§ 언어모델

§ 멀티모달모델

§ 데이터처리시스템§ 배치처리

§ 스트림처리

§ 자동최적화

AI Systems

Serving

ML Framework (Compiler, Runtime)

Cluster Manager

Hardware

Training

WorkflowAutoML

SystemY

NimbleTerra

WindTunnel

Band

SystemX

Terra: Imperative-Symbolic Co-Execution of Imperative Deep Learning Programs (NeurIPS 2021)

Nimble: Lightweight and Parallel GPU Task Scheduling for Deep Learning (NeurIPS 2020 Spotlight)

Graph

Pre-run (base framework)

GPU KernelDispatching

GPU MemoryRequest

Prepare KernelFunction Args.

GPU Task Submission

Task Schedule

AoT GPU Task Scheduler

Load Inputs to GPU MemoryRuntime

Replay Recorded Tasks

…

…

Execution TraceReserved Memory GPU task A

GPU task BGPU task C…

Input Tensors

https://github.com/snuspl/nimble

WindTunnel: Towards Differentiable ML Pipelines Beyond a Single Model (VLDB 2022)

Parallax: Sparsity-aware Distributed Training (EuroSys 2019) & Elastic Parallax

10

Machine

GPU GPU

Machine

GPU GPU

2 1 5 2 1 5 2 1 5 2 1 5

Server Server

GradAgg(AR)

GradAgg(PS)

GPU GradAgg(PS)

Local Agg

Dense var.

Sparse var.

Local Agg

https://github.com/snuspl/parallax

Crane: A Declarative GPU Cluster Manager

Apache Nemo (ATC 2019, TOCS 2021, Apache Incubator Project)

.map(…)

.reduce(…)

RuntimeMaster

Frontend Optimizer

Backend

Compiler

BackboneModules

Optimization pass:IR à IR’

Optimization policy

User application

Executor

BackboneModules

Intermediaterepresentation (IR)

Optimized IR

https://nemo.apache.orghttps://github.com/apache/incubator-nemo

Blaze & Sponge

BlazeRunner

Spark App

ProfilingPhase

(1)

Shared disaggregated storage memory

SparkExecutor

EM

Spark Master

rdd1_1 rdd1_2 rdd2_1rdd1_1rdd1_1 rdd1_2rdd1_2 rdd2_1rdd2_1

SparkExecutor SparkExecutor

UtilityFunction

Caching/EvictionPolicy

caching/eviction on

disaggregated SM

SM EM SM EM SM

Application DAG(2) Execution

Phase(4)

(3) caching/eviction

on local SM

(2)

Inter-job Caching for Iterative Data Processing Fast Stream Processing Scaling on Lambda

Blaze Sponge

Fluffy: Finding Consensus Bugs in Ethereum via Multi-transaction Differential Fuzzing (OSDI 2021)

Found two new consensus bugs in Ethereumvia a new fuzzing technique

https://github.com/snuspl/fluffy

§ Neural Architecture Search

§ Hyper-Parameter Optimization

§ Inter-job Caching

§ Stream Processing on Serverless Frameworks

§ Mobile Deep Learning Runtime

§ …

Under Submission

서울대-네이버하이퍼스케일 AI 센터

FriendliAI: 초거대 AI 개발플랫폼스타트업

https://friendli.ai (2022년 병역특례업체신청예정)

서울대학교

Thank you!

bgchun (at) snu.ac.krbgchun (at) friendli.ai

Software Platform Lab 소개