Rui's Blog
  • Rui's Blog/Paper Reading Notes - Introduction
  • Personal Blog
    • Personal Blog - Index
      • How to Create Picture-in-Picture Effect / Video Overlay for a Presentation Video
      • How to Do Your Part to Protect the Environment in Wisconsin
      • How to Get a Driver's License in Wisconsin
      • How to Travel from the U.S. to China onboard AA127 in June 2021
      • How to Transfer Credits Back to UW-Madison
      • Resources on Learning Academic Writing (for Computer Science)
    • Towards applying to CS Ph.D. programs
  • Machine Learning Systems
    • Machine Learning Systems - Index
      • MLSys Papers - Short Notes
      • [2011 NSDI] Dominant Resource Fairness: Fair Allocation of Multiple Resource Types
      • [2014 OSDI] Scaling Distributed Machine Learning with the Parameter Server
      • [2018 OSDI] Gandiva: Introspective Cluster Scheduling for Deep Learning
      • [2018 SIGCOMM] Chameleon: Scalable Adaptation of Video Analytics via Temporal and Cross-camera ...
      • [2018 NIPS] Dynamic Space-Time Scheduling for GPU Inference
      • [2019 ATC] Analysis of Large-Scale Multi-Tenant GPU Clusters for DNN Training Workloads
      • [2019 NSDI] Tiresias: A GPU Cluster Manager for Distributed Deep Learning
      • [2019 SOSP] ByteScheduler: A Generic Communication Scheduler for Distributed DNN Training ...
      • [2019 SOSP] PipeDream: Generalized Pipeline Parallelism for DNN Training
      • [2019 SOSP] Parity Models: Erasure-Coded Resilience for Prediction Serving Systems
      • [2019 NIPS] GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism
      • [2019 SC] ZeRO: memory optimizations toward training trillion parameter models
      • [2020 OSDI] Gavel: Heterogeneity-Aware Cluster Scheduling Policies for Deep Learning Workloads
      • [2020 OSDI] AntMan: Dynamic Scaling on GPU Clusters for Deep Learning
      • [2020 OSDI] BytePS: A High Performance and Generic Framework for Distributed DNN Training
      • [2020 SIGCOMM] Reducto: On-Camera Filtering for Resource-Efficient Real-Time Video Analytics
        • [2020 MLSys] Salus: Fine-Grained GPU Sharing Primitives for Deep Learning Applications
      • [2020 EuroSys] AlloX: Compute Allocation in Hybrid Clusters
      • [2020 VLDB] PyTorch Distributed: Experiences on Accelerating Data Parallel Training
      • [2020 NetAI] Is Network the Bottleneck of Distributed Training?
      • [2020 NSDI] Themis: Fair and Efficient GPU Cluster Scheduling
      • [2021 MLSys] Accordion: Adaptive Gradient Communication via Critical Learning Regime Identification
      • [2021 VLDB] Analyzing and Mitigating Data Stalls in DNN Training
      • [2021 FAST] CheckFreq: Frequent, Fine-Grained DNN Checkpointing
      • [2021 EuroMLSys] Interference-Aware Scheduling for Inference Serving
      • [2021 OSDI] Pollux: Co-adaptive Cluster Scheduling for Goodput-Optimized Deep Learning
      • [2021 MLSys] Wavelet: Efficient DNN Training with Tick-Tock Scheduling
      • [2021 NSDI] SwitchML: Scaling Distributed Machine Learning with In-Network Aggregation
    • Big Data Systems - Index
      • Big Data Systems Papers - Short Notes
      • [2003 SOSP] The Google File System
      • [2004 OSDI] MapReduce: Simplified Data Processing on Large Clusters
      • [2010 SIGMOD] Pregel: A System for Large-Scale Graph Processing
      • [2011 NSDI] Mesos: A Platform for Fine-Grained Resource Sharing in the Data Center
      • [2012 NSDI] Resilient Distributed Datasets: A Fault-Tolerant Abstraction for In-Memory Cluster ...
      • [2012 OSDI] PowerGraph: Distributed Graph-Parallel Computation on Natural Graphs
      • [2019 FAST] DistCache: Provable Load Balancing for Large-Scale Storage Systems with Distributed...
      • [2021 HotOS] From Cloud Computing to Sky Computing
      • [2021 EuroSys] NextDoor: Accelerating graph sampling for graph machine learning using GPUs
  • Earlier Readings & Notes
    • High Performance Computing Course Notes
      • Lecture 1: Course Overview
      • Lecture 2: From Code to Instructions. The FDX Cycle. Instruction Level Parallelism.
      • Lecture 3: Superscalar architectures. Measuring Computer Performance. Memory Aspects.
      • Lecture 4: The memory hierarchy. Caches.
      • Lecture 5: Caches, wrap up. Virtual Memory.
      • Lecture 6: The Walls to Sequential Computing. Moore’s Law.
      • Lecture 7: Parallel Computing. Flynn's Taxonomy. Amdahl's Law.
      • Lecture 8: GPU Computing Intro. The CUDA Programming Model. CUDA Execution Configuration.
      • Lecture 9: GPU Memory Spaces
      • Lecture 10: GPU Scheduling Issues.
      • Lecture 11: Execution Divergence. Control Flow in CUDA. CUDA Shared Memory Issues.
      • Lecture 12: Global Memory Access Patterns and Implications.
      • Lecture 13: Atomic operations in CUDA. GPU ode optimization rules of thumb.
      • Lecture 14: CUDA Case Studies. (1) 1D Stencil Operation. (2) Vector Reduction in CUDA.
      • Lecture 15: CUDA Case Studies. (3) Parallel Prefix Scan on the GPU. Using Multiple Streams in CUDA.
      • Lecture 16: Streams, and overlapping data copy with execution.
      • Lecture 17: GPU Computing: Advanced Features.
      • Lecture 18: GPU Computing with thrust and cub.
      • Lecture 19: Hardware aspects relevant in multi-core, shared memory parallel computing.
      • Lecture 20: Multi-core Parallel Computing with OpenMP. Parallel Regions.
      • Lecture 21: OpenMP Work Sharing.
      • Lecture 22: OpenMP Work Sharing
      • Lecture 23: OpenMP NUMA Aspects. Caching and OpenMP.
      • Lecture 24: Critical Thinking. Code Optimization Aspects.
      • Lecture 25: Computing with Supercomputers.
      • Lecture 26: MPI Parallel Programming General Introduction. Point-to-Point Communication.
      • Lecture 27: MPI Parallel Programming Point-to-Point communication: Blocking vs. Non-blocking sends.
      • Lecture 28: MPI Parallel Programming: MPI Collectives. Overview of topics covered in the class.
    • Cloud Computing Course Notes
      • 1.1 Introduction to Clouds, MapReduce
      • 1.2 Gossip, Membership, and Grids
      • 1.3 P2P Systems
      • 1.4 Key-Value Stores, Time, and Ordering
      • 1.5 Classical Distributed Algorithms
      • 4.1 Spark, Hortonworks, HDFS, CAP
      • 4.2 Large Scale Data Storage
    • Operating Systems Papers - Index
      • CS 736 @ UW-Madison Fall 2020 Reading List
      • All File Systems Are Not Created Equal: On the Complexity of Crafting Crash-Consistent Applications
      • ARC: A Self-Tuning, Low Overhead Replacement Cache
      • A File is Not a File: Understanding the I/O Behavior of Apple Desktop Applications
      • Biscuit: The benefits and costs of writing a POSIX kernel in a high-level language
      • Data Domain: Avoiding the Disk Bottleneck in the Data Domain Deduplication File System
      • Disco: Running Commodity Operating Systems on Scalable Multiprocessors
      • FFS: A Fast File System for UNIX
      • From WiscKey to Bourbon: A Learned Index for Log-Structured Merge Trees
      • LegoOS: A Disseminated, Distributed OS for Hardware Resource Disaggregation
      • LFS: The Design and Implementation of a Log-Structured File System
      • Lottery Scheduling: Flexible Proportional-Share Resource Management
      • Memory Resource Management in VMware ESX Server
      • Monotasks: Architecting for Performance Clarity in Data Analytics Frameworks
      • NFS: Sun's Network File System
      • OptFS: Optimistic Crash Consistency
      • RAID: A Case for Redundant Arrays of Inexpensive Disks
      • RDP: Row-Diagonal Parity for Double Disk Failure Correction
      • Resource Containers: A New Facility for Resource Management in Server Systems
      • ReVirt: Enabling Intrusion Analysis through Virtual-Machine Logging and Replay
      • Scheduler Activations: Effective Kernel Support for the User-Level Management of Parallelism
      • SnapMirror: File-System-Based Asynchronous Mirroring for Disaster Recovery
      • The Linux Scheduler: a Decade of Wasted Cores
      • The Unwritten Contract of Solid State Drives
      • Venti: A New Approach to Archival Storage
    • Earlier Notes
      • How to read a paper
  • FIXME
    • Template for Paper Reading Notes
Powered by GitBook
On this page
  • One-line Summary
  • Paper Structure Outline
  • Background & Motivation
  • Design and Implementation
  • Data parallelism
  • Model parallelism
  • Evaluation
  • Data parallelism
  • Model parallelism
  • Overhead analysis
  • Links

Was this helpful?

  1. Machine Learning Systems
  2. Machine Learning Systems - Index

[2021 MLSys] Wavelet: Efficient DNN Training with Tick-Tock Scheduling

Previous[2021 OSDI] Pollux: Co-adaptive Cluster Scheduling for Goodput-Optimized Deep LearningNext[2021 NSDI] SwitchML: Scaling Distributed Machine Learning with In-Network Aggregation

Last updated 2 years ago

Was this helpful?

One-line Summary

Both data and model parallelism suffer from system under-utilization. Wavelet exploits the under-utilized memory & compute by scaling up the number of training tasks and launching the additional tasks with a delay to fully utilize the on-chip memory and improve the compute utilization, speeding up individual jobs.

That was an unnecessarily long sentence... GRE took its toll on me!

Paper Structure Outline

  1. Introduction

  2. Background and Motivation

    1. Distributed DNN Training Schemes

    2. Jobs Characteristics of Distributed DNN Training

      1. Zoom-in analysis on data parallel training

      2. Sub-iteration analysis on model parallel training

  3. Wavelet Design

    1. System Overview

    2. Wavelet in Data Parallelism

      1. Memory overlapping

      2. Computation overlapping

      3. Model synchronization between waves

    3. Wavelet in Model Parallelism

      1. Launching multiple tock-wave tasks

      2. Model partition switching

      3. Inter-batch synchronization

  4. Evaluation

    1. Data parallelism

      1. Single machine multi-GPU

      2. Multi-machine multi-GPU

    2. Model parallelism

      1. Single machine multi-GPU

      2. Multi-machine multi-GPU

      3. Overhead analysis

  5. Related Work

    1. Resource allocation for distributed DNN training

    2. GPU sharing

  6. Conclusion

Background & Motivation

Bigger models & datasets call for large-scale distributed machine learning training. The current scheduling policy, gang scheduling, where all training tasks on all workers need to be launched at the same time, contributes to the under-utilization of system resources (compute & memory).

In Fig. 1 (see above), computation is memory-bounded during the forward propagation. Between time 0.4 and 0.6, memory is underutilized in the backward propagation. Moreover, ~60% of on-chip compute cores are underutilized.

Design and Implementation

Data parallelism

Model synchronization

In vanilla allreduce, there are only Main & Tick waves. With this Tock wave added, Wavelet doubles the number of model synchronizations. This is the same as synchronizing over 2*N data parallel tasks on 2*N GPUs, thus guaranteeing convergence.

Overlapping memory

In gang scheduling, the memory of all GPUs is underutilized during backprop. Tick-tock scheduling injects tock-wave tasks right after the tick-wave tasks finish the forward pass. To concurrently run 2 tasks (tick & tock), 2 model replicas are maintained on the GPU since the two waves train on different data. In the memory, the size of the model is way smaller than the size of the intermediate results, so no need to worry about the extra memory.

Overlapping computation

CUDA computation kernels are launched in separate CUDA streams to ensure ordered execution within a stream and non-blocking across different streams. The empty bubbles between kernels is due to the latency of CPUs sending instructions to GPUs.

Model parallelism

In the vanilla pipelined process (white blocks), only 1 batch is active in the system and at most 1 GPU is active at a time. Each GPU also holds the same model partition during the whole training process.

In Wavelet, we inject 3 (N-1 w/ N GPUs) tock waves on 1 tick wave. The model partition is swapped on each GPU using a round-robin fashion. There exists an extra model synchronization for each model partition, and the context switching also brings overhead.

Evaluation

Data parallelism

  • Single machine: Up to 1.88x speedup (avg: 1.4x, theoretically 2x) over DP baseline

  • Multiple machine: Up to 1.18x speedup over baseline. The worse throughput than baseline is the overhead kicking in: The cross-machine low-bandwidth network becomes the bottleneck during the extra allreduce

Model parallelism

  • Only ~2.5x speedup in 4x/8x parallelism

    • Gpipe/PipeDream breaks a mini-batch into smaller micro-batches -> High-frequency but small-size data chunks

    • Number of CUDA kernel calls ↑

    • Intermediate result transfer between GPUs that hold different model partitions ↑

    • Linear scalability in theory

Overhead analysis

  • Context switch: Switching model partition, ~4% of total training time

  • Communication: Transferring intermediate results across GPUs (~15% of total training time)

  • All reduce: Model synchronization during backprop (~4% of total training time)

Links

There are existing job multiplexing schemes that boost system utilization. lets 2 low-utilization jobs space-share a GPU. provides fine-grained memory sharing via the GPU Lane abstraction. However, neither scheme contributes to the training progress of a single job. In this work, Wavelet relaxes the gang scheduling scheme and accelerates a single job while improving the system utilization.

Note that this is in log-scale

Gandiva
Salus
Paper PDF
Presentation video at MLSys '21
Presentation slides at MLSys '21
Example of how Wavelet is applied to data parallel training
The same thing happens for model parallelism where the memory valley is longer. With pipelining = using GPipe.
How Wavelet handles the extra tock-wave during model synchronization