본문 바로가기 주메뉴 바로가기
22대 대한민국 국회 국회정보길라잡이 국회의원검색
회원가입 로그인
HOME

정보검색

소장정보 검색

국회도서관 홈으로 정보검색 소장정보 검색

결과 내 검색

동의어 포함

고급검색

인명/단체명

저자정보 상세정보
인명/단체명을 입력하세요.

키워드

  • 대표어

  • 외국어

-

목차보기

Title Page

Abstract

Contents

I. Introduction 9

1.1. Contribution 11

1.2. Organization 11

II. Background 13

2.1. Deep Neural Network Training 13

2.2. GPU Sharing Use Cases 13

2.3. Spatial GPU Sharing 14

III. Challenges for Memory Sharing 15

3.1. Memory Bloating 15

3.2. Workload Variability 16

3.3. Asynchrony with GPU Processing 16

IV. Design Overview 18

V. Scheduling Algorithm 21

5.1. Problem Definition 21

5.2. Time Shift Model 21

5.3. Memory Sharing Algorithm 22

VI. Memory Management with Concurrency 25

6.1. Tracking Memory Usage in GPU 25

6.2. Tensor Classification 26

6.3. Managing Memory Regions 26

VII. Evaluation 28

7.1. Training Same Models 28

7.2. Training Non-identical Models 30

7.3. Dynamic Memory Budget Change 31

7.4. Design Validation 32

VIII. Discussion 34

IX. Related Work 35

X. Concluding Remarks 36

References 37

List of Figures

Figure 1. Cumulative distribution of NASNet tensor lifespan. 15

Figure 2. Memory usage patterns for different DNN models over time. 15

Figure 3. GPU memory usage from CPU view (ResNet-50). 17

Figure 4. System architecture in Zico. 18

Figure 5. Throughput in training the same models.... 28

Figure 6. Aggregated memory usage for training the same models.... 29

Figure 7. Memory usage over time for training the same models.... 29

Figure 8. Throughput in training the distinct models concurrently.... 30

Figure 9. Memory usage over time for training NASNet and ResNet-110 concurrently. 30

Figure 10. Memory usage patterns on dynamic memory budget changes. 31

초록보기

GPUs are the workhorse in modern server infrastructure fueling advances in a number of compute-intensive workloads such as deep neural network (DNN) training. Several recent works propose solutions on sharing GPU resources across multiple concurrent DNN training jobs, but none of them address rapidly increasing memory footprint introduced by such job co-locations, which greatly limit the effectiveness of sharing GPU resources. In this paper, Zico, the first DNN system, is presented which aims at reducing the system-wide memory consumption for concurrent training. Zico keeps track of the memory usage pattern of individual training job by monitoring its progress on GPU computations and makes memory reclaimed from the job globally sharable. Based on this memory management scheme, Zico automatically decides a strategy to share memory among concurrent jobs with minimum delay on training while not exceeding a given memory budget such as GPU memory capacity. Our evaluation shows that Zico outperforms existing GPU sharing approaches and delivers benefits over a variety of job co-location scenarios.

추천서가 (다양한 추천 자료를 만나보세요)

권호기사보기

권호기사 목록 테이블로 기사명, 저자명, 페이지, 원문, 기사목차 순으로 되어있습니다.
기사명 저자명 페이지 원문 기사목차

권호

기사명 원문보기 기사목차
닫기