Large-scale model training has been a playing ground for a limited few requiring complex model refactoring and access to prohibitively expensive GPU clusters. ZeRO-Offload changes the large model training landscape by making large model training accessible to nearly everyone. It can train models with over 13 billion parameters on a single GPU, a 10x increase in size compared to popular framework such as PyTorch, and it does so without requiring any model change from the data scientists or sacrificing computational efficiency. ZeRO-Offload enables large model training by offloading data and compute to CPU. To preserve compute efficiency, it is designed to minimize the data movement to/from GPU, and reduce CPU compute time while maximizing memory savings on GPU. As a result, ZeRO-Offload can achieve 40 TFlops/GPU on a single NVIDIA V100 GPU for 10B parameter model compared to 30TF using PyTorch alone for a 1.4B parameter model, the largest that can be trained without running out of memory. ZeRO-Offload is also designed to scale on multiple-GPUs when available, offering near linear speedup on up to 128 GPUs. Additionally, it can work together with model parallelism to train models with over 70 billion parameters on a single DGX-2 box, a 4.5x increase in model size compared to using model parallelism alone. By combining compute and memory efficiency with ease-of-use, ZeRO-Offload democratizes large-scale model training making it accessible to even data scientists with access to just a single GPU.

ZeRO-Offload: Democratizing Billion-Scale Model Training

In the last three years, the largest dense deep learning models have grown over 1000x to reach hundreds of billions of parameters, while the GPU memory has only grown by 5x (16 GB to 80 GB). Therefore, the growth in model scale has been supported primarily though system innovations that allow large models to fit in the aggregate GPU memory of multiple GPUs. However, we are getting close to the GPU memory wall. It requires 800 NVIDIA V100 GPUs just to fit a trillion parameter model for training, and such clusters are simply out of reach for most data scientists. In addition, training models at that scale requires complex combinations of parallelism techniques that puts a big burden on the data scientists to refactor their model. In this paper we present ZeRO-Infinity, a novel heterogeneous system technology that leverages GPU, CPU, and NVMe memory to allow for unprecedented model scale on limited resources without requiring model code refactoring. At the same time it achieves excellent training throughput and scalability, unencumbered by the limited CPU or NVMe bandwidth. ZeRO-Infinity can fit models with tens and even hundreds of trillions of parameters for training on current generation GPU clusters. It can be used to fine-tune trillion parameter models on a single NVIDIA DGX-2 node, making large models more accessible. In terms of training throughput and scalability, it sustains over 25 petaflops on 512 NVIDIA V100 GPUs (40% of peak), while also demonstrating super linear scalability. An open source implementation of ZeRO-Infinity is available through DeepSpeed 1.

ZeRO-infinity: breaking the GPU memory wall for extreme scale deep learning

In the last three years, the largest dense deep learning models have grown over 1000x to reach hundreds of billions of parameters, while the GPU memory has only grown by 5x (16 GB to 80 GB). Therefore, the growth in model scale has been supported primarily though system innovations that allow large models to fit in the aggregate GPU memory of multiple GPUs. However, we are getting close to the GPU memory wall. It requires 800 NVIDIA V100 GPUs just to fit a trillion parameter model for training, and such clusters are simply out of reach for most data scientists. In addition, training models at that scale requires complex combinations of parallelism techniques that puts a big burden on the data scientists to refactor their model. 
In this paper we present ZeRO-Infinity, a novel heterogeneous system technology that leverages GPU, CPU, and NVMe memory to allow for unprecedented model scale on limited resources without requiring model code refactoring. At the same time it achieves excellent training throughput and scalability, unencumbered by the limited CPU or NVMe bandwidth. ZeRO-Infinity can fit models with tens and even hundreds of trillions of parameters for training on current generation GPU clusters. It can be used to fine-tune trillion parameter models on a single NVIDIA DGX-2 node, making large models more accessible. In terms of training throughput and scalability, it sustains over 25 petaflops on 512 NVIDIA V100 GPUs(40% of peak), while also demonstrating super linear scalability. An open source implementation of ZeRO-Infinity is available through DeepSpeed, a deep learning optimization library that makes distributed training easy, efficient, and effective.

ZeRO-Infinity: Breaking the GPU Memory Wall for Extreme Scale Deep Learning.

Molecular dynamics (MD) simulation is a fundamental method for modeling ensembles of particles. In this paper, we introduce a new method to improve the performance of MD by leveraging the emerging TB-scale big memory system. In particular, we trade memory capacity for computation capability to improve MD performance by the lookup table-based memoization technique. The traditional memoization technique for the MD simulation uses relatively small DRAM, bases on a suboptimal data structure, and replaces pair-wise computation, which leads to limited performance benefit in the big memory system. We introduce MD-HM, a memoization-based MD simulation framework customized for the big memory system. MD-HM partitions the simulation field into subgrids, and replaces computation in each subgrid as a whole based on a lightweight pattern-match algorithm to recognize computation in the subgrid. MD-HM uses a new two-phase LSM-tree to optimize read/write performance. Evaluating with nine MD simulations, we show that MD-HM outperforms the state-of-the-art LAMMPS simulation framework with an average speedup of 7.6x based on the Intel Optane-based big memory system.

https://dl.acm.org/doi/pdf/10.1145/3447818.3460365

MD-HM: memoization-based molecular dynamics simulations on big memory system

Graph neural networks (GNNs) can extract features by learning both the representation of each objects (i.e., graph nodes) and the relationship across different objects (i.e., the edges that connect nodes), achieving state-of-the-art performance in various graph-based tasks. Despite its strengths, utilizing these algorithms in a production environment faces several challenges as the number of graph nodes and edges amount to several billions to hundreds of billions scale, requiring substantial storage space for training. Unfortunately, state-of-the-art ML frameworks employ an in-memory processing model which significantly hampers the productivity of ML practitioners as it mandates the overall working set to fit within DRAM capacity. In this work, we first conduct a detailed characterization on a state-of-the-art, large-scale GNN training algorithm, GraphSAGE. Based on the characterization, we then explore the feasibility of utilizing capacity-optimized NVMe SSDs for storing memory-hungry GNN data, which enables large-scale GNN training beyond the limits of main memory size. Given the large performance gap between DRAM and SSD, however, blindly utilizing SSDs as a direct substitute for DRAM leads to significant performance loss. We therefore develop SmartSAGE, our software/hardware co-design based on an in-storage processing (ISP) architecture. Our work demonstrates that an ISP based large-scale GNN training system can achieve both high capacity storage and high performance, opening up opportunities for ML practitioners to train large GNN datasets without being hampered by the physical limitations of main memory size.

SmartSAGE: training large-scale graph neural networks using in-storage processing architectures

Memory capacity is a major bottleneck for training deep neural networks (DNN). Heterogeneous memory (HM) combining fast and slow memories provides a promising direction to increase memory capacity. However, HM imposes challenges on tensor migration and allocation for high performance DNN training. Prior work heavily relies on DNN domain knowledge, unnecessarily causes tensor migration due to page-level false sharing, and wastes fast memory space. We present Sentinel, a software runtime system that automatically optimizes tensor management on HM. Sentinel uses dynamic profiling, and coordinates operating system (OS) and runtime-level profiling to bridge the semantic gap between OS and applications, which enables tensor-level profiling. This profiling enables co-allocating tensors with similar lifetime and memory access frequency into the same pages. Such fine-grained profiling and tensor collocation avoids unnecessary data movement, improves tensor movement efficiency, and enables larger batch training because of saving in fast memory space. Sentinel reduces fast memory consumption by 80% while retaining comparable performance to fast memory-only system; Sentinel consistently outperforms a state-of-the-art solution on CPU by 37% and two state-of-the-art solutions on GPU by 2x and 21% respectively in training throughput.

Sentinel: Efficient Tensor Migration and Allocation on Heterogeneous Memory Systems for Deep Learning

Particle simulations of plasma are important for understanding plasma dynamics in space weather and fusion devices. However, production simulations that use billions and even trillions of computational particles require high memory capacity. In this work, we explore the latest persistent memory (PM) hardware to enable large-scale plasma simulations at unprecedented scales on a single machine. We use WarpX, an advanced plasma simulation code which is mission-critical and targets future exascale systems. We analyze the performance of WarpX on PM-based heterogeneous memory systems and propose to make the best use of memory hierarchy to avoid the impact of inferior performance of PM. We introduce a combination of static and dynamic data placement, and processor-cache prefetch mechanism for performance optimization. We develop a performance model to enable efficient data migration between PM and DRAM in the background, without reducing available bandwidth and parallelism to the application threads. We also build an analytical model to decide when to prefetch for the best use of caches. Our design achieves 66.4% performance improvement over the PM-only baseline and outperforms DRAM-cached, NUMA first-touch, and a state-of-the-art software solution by 38.8%, 45.1% and 83.3%, respectively.

https://dl.acm.org/doi/pdf/10.1145/3447818.3460356

Optimizing large-scale plasma simulations on persistent memory-based heterogeneous memory with effective data placement across memory hierarchy

Software-managed heterogeneous memory (HM) provides a promising solution to increase memory capacity and cost efficiency. However, to release the performance potential of HM, we face a problem of data management. Given an application with various execution phases and each with possibly distinct working sets, we must move data between memory components of HM to optimize performance. The deep neural network (DNN), as a common workload on data centers, imposes great challenges on data management on HM. This workload often employs a task dataflow execution model, and is featured with a large amount of small data objects and fine-grained operations (tasks). This execution model imposes challenges on memory profiling and efficient data migration. 
We present Sentinel, a runtime system that automatically optimizes data migration (i.e., data management) on HM to achieve performance similar to that on the fast memory-only system with a much smaller capacity of fast memory. To achieve this,Sentinel exploits domain knowledge about deep learning to adopt a custom approach for data management. Sentinel leverages workload repeatability to break the dilemma between profiling accuracy and overhead; It enables profiling and data migration at the granularity of data objects (not pages), by controlling memory allocation. This method bridges the semantic gap between operating system and applications. By associating data objects with the DNN topology, Sentinel avoids unnecessary data movement and proactively triggers data movement. Using only 20% of peak memory consumption of DNN models as fast memory size, Sentinel achieves the same or comparable performance (at most 8% performance difference) to that of the fast memory-only system on common DNN models; Sentinel also consistently outperforms a state-of-the-art solution by 18%.

Jiaolin Luo

Papers

Sentinel: Efficient Tensor Migration and Allocation on Heterogeneous Memory Systems for Deep Learning

Optimizing large-scale plasma simulations on persistent memory-based heterogeneous memory with effective data placement across memory hierarchy

Sentinel: Runtime Data Management on Heterogeneous Main MemorySystems for Deep Learning.