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
  • Summary
  • Background & Motivation
  • Design & Implementation
  • ZeRO-DP Stage 1: Optimizer State Partitioning
  • ZeRO-DP Stage 2: Gradient Partitioning
  • ZeRO-DP Stage 3: Parameter Partitioning
  • ZeRO-R
  • Evaluation
  • ZeRO-Infinity and ZeRO-Offload
  • Links & References

Was this helpful?

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

[2019 SC] ZeRO: memory optimizations toward training trillion parameter models

Previous[2019 NIPS] GPipe: Efficient Training of Giant Neural Networks using Pipeline ParallelismNext[2020 OSDI] Gavel: Heterogeneity-Aware Cluster Scheduling Policies for Deep Learning Workloads

Last updated 2 years ago

Was this helpful?

Summary

ZeRO is a new distributed training paradigm that vastly improves the memory efficiency of large-scale model training.

Background & Motivation

Existing solutions for distributed training include data parallelism (DP), model parallelism (MP), pipeline parallelism (PP), 3D parallelism, CPU offloading, etc., each with a couple of catches:

  • DP has good compute/communication efficiency but poor memory efficiency. In DP, each parallel worker holds a full copy of the model, so each device quickly runs out of memory (e.g., for models with > 1.4B parameters on a GPU with 32 GB GRAM).

  • MP has good memory efficiency but poor compute/communication efficiency. It splits the model vertically, requiring significant communications between each layer on different devices. As a result of that, the efficiency quickly degrades when devices become far apart from each other: for example, when training a 40B-parameter model across two DGX nodes, each GPU's computing efficiency (tflops) is only 5% of the hardware peak.

Is there a way to achieve the best of all worlds? The Microsoft folks first take a look at the spectrum of memory consumption in large-model training, and they classify the memory consumption into two parts:

  1. Model states: parameters, gradients, and optimizer states (e.g., momentum and variances in Adam). These take the majority of the memory in large-model training.

  2. Residual states: activations, temporary buffers, and unusable fragmented memory.

The authors develop ZeRO-DP to address (1) and ZeRO-R to address (2).

Design & Implementation

A key insight of ZeRO-DP is that both DP and MP keep all the model states needed over the entire training process, but not everything is required all the time. For example, parameters corresponding to each layer are only needed during the forward/backward propagation of the layer.

ZeRO-DP Stage 1: Optimizer State Partitioning

The optimizer states are grouped into N equal partitions (N is the DP degree), such that worker i only stores and updates the optimizer states of partition i (1/N of the total optimizer states and parameters).

ZeRO-DP Stage 2: Gradient Partitioning

As each DP process only updates its corresponding parameter partition, it only needs the reduced gradients for the corresponding parameters. Therefore, as each gradient of each layer becomes available during the backward propagation, we only reduce them on the DP process responsible for updating the corresponding parameters. After the reduction, we no longer need the gradients and their memory can be released. The total communication volume is the same as vanilla DP (all-reduce = reduce-scatter + all-gather), because in this setup, a reduce-scatter is performed on the gradients and an all-gather on the parameters.

ZeRO-DP Stage 3: Parameter Partitioning

In this stage, each process only stores the parameters corresponding to its partition. When the parameters outside of its partition are required for forward/backward propagation, they are received from the appropriate DP process through broadcast. This technique only increases the total communication volume to 1.5x of a DP baseline (two all-gathers of the parameters are required for forward/back prop, and the gradients need to be reduce-scattered), while reducing the per-worker memory consumption by N times. With all these optimizations, trillion-parameter models can be trained on thousands of modern-day GPUs.

ZeRO-R

A lot of follow-up works of ZeRO focuses on ZeRO-DP, so I'll come back to read this section later

Evaluation

ZeRO-Infinity and ZeRO-Offload

ZeRO-Infinity and ZeRO-Offload are follow-up systems that offload data and compute to CPUs and NVMe.

Links & References

Different implementations of PP have different issues. For example, requires a batch size proportional to the number of pipeline partitions, and a large batch size might hurt convergence. is very memory inefficient due to weight stashing.

Blog: . This includes a nice video that explains how ZeRO-DP works.

G-pipe
PipeDream
Paper PDF
Presentation video at a webinar
ZeRO & DeepSpeed: New system optimizations enable training models with over 100 billion parameters
DeepSpeed on GitHub