Vijay Chidambaram

Aashaka Shah, Chao-Yuan Wu, Jayashree Mohan et al.

Deep learning is slowly, but steadily, hitting a memory bottleneck. While the tensor computation in top-of-the-line GPUs increased by 32x over the last five years, the total available memory only grew by 2.5x. This prevents researchers from exploring larger architectures, as training large networks requires more memory for storing intermediate outputs. In this paper, we present MONeT, an automatic framework that minimizes both the memory footprint and computational overhead of deep networks. MONeT jointly optimizes the checkpointing schedule and the implementation of various operators. MONeT is able to outperform all prior hand-tuned operations as well as automated checkpointing. MONeT reduces the overall memory requirement by 3x for various PyTorch models, with a 9-16% overhead in computation. For the same computation cost, MONeT requires 1.2-1.8x less memory than current state-of-the-art automated checkpointing frameworks. Our code is available at https://github.com/utsaslab/MONeT.

Aashaka Shah, Vijay Chidambaram, Meghan Cowan et al.

Machine learning models are increasingly being trained across multiple GPUs and servers. In this setting, data is transferred between GPUs using communication collectives such as AlltoAll and AllReduce, which can become a significant bottleneck in training large models. Thus, it is important to use efficient algorithms for collective communication. We develop TACCL, a tool that enables algorithm designers to guide a synthesizer into automatically generating algorithms for a given hardware configuration and communication collective. TACCL uses a novel communication sketch abstraction to get crucial information from the designer to significantly reduce the search space and guide the synthesizer towards better algorithms. TACCL also uses a novel encoding of the problem that allows it to scale beyond single-node topologies. We use TACCL to synthesize algorithms for three collectives and two hardware topologies: DGX-2 and NDv2. We demonstrate that the algorithms synthesized by TACCL outperform the Nvidia Collective Communication Library (NCCL) by up to 6.7x. We also show that TACCL can speed up end-to-end training of Transformer-XL and BERT models by 11%--2.3x for different batch sizes.

Jayashree Mohan, Amar Phanishayee, Janardhan Kulkarni et al.

Training Deep Neural Networks (DNNs) is a widely popular workload in both enterprises and cloud data centers. Existing schedulers for DNN training consider GPU as the dominant resource, and allocate other resources such as CPU and memory proportional to the number of GPUs requested by the job. Unfortunately, these schedulers do not consider the impact of a job's sensitivity to allocation of CPU, memory, and storage resources. In this work, we propose Synergy, a resource-sensitive scheduler for shared GPU clusters. Synergy infers the sensitivity of DNNs to different resources using optimistic profiling; some jobs might benefit from more than the GPU-proportional allocation and some jobs might not be affected by less than GPU-proportional allocation. Synergy performs such multi-resource workload-aware assignments across a set of jobs scheduled on shared multi-tenant clusters using a new near-optimal online algorithm. Our experiments show that workload-aware CPU and memory allocations can improve average JCT up to 3.4x when compared to traditional GPU-proportional scheduling.

Zsolt Istvan, Soujanya Ponnapalli, Vijay Chidambaram

Enforcing data protection and privacy rules within large data processing applications is becoming increasingly important, especially in the light of GDPR and similar regulatory frameworks. Most modern data processing happens on top of a distributed storage layer, and securing this layer against accidental or malicious misuse is crucial to ensuring global privacy guarantees. However, the performance overhead and the additional complexity for this is often assumed to be significant -- in this work we describe a path forward that tackles both challenges. We propose "Software-Defined Data Protection" (SDP), an adoption of the "Software-Defined Storage" approach to non-performance aspects: a trusted controller translates company and application-specific policies to a set of rules deployed on the storage nodes. These, in turn, apply the rules at line-rate but do not take any decisions on their own. Such an approach decouples often changing policies from request-level enforcement and allows storage nodes to implement the latter more efficiently. Even though in-storage processing brings challenges, mainly because it can jeopardize line-rate processing, we argue that today's Smart Storage solutions can already implement the required functionality, thanks to the separation of concerns introduced by SDP. We highlight the challenges that remain, especially that of trusting the storage nodes. These need to be tackled before we can reach widespread adoption in cloud environments.

Jayashree Mohan, Amar Phanishayee, Ashish Raniwala et al.

Training Deep Neural Networks (DNNs) is resource-intensive and time-consuming. While prior research has explored many different ways of reducing DNN training time, the impact of input data pipeline, i.e., fetching raw data items from storage and performing data pre-processing in memory, has been relatively unexplored. This paper makes the following contributions: (1) We present the first comprehensive analysis of how the input data pipeline affects the training time of widely-used computer vision and audio Deep Neural Networks (DNNs), that typically involve complex data preprocessing. We analyze nine different models across three tasks and four datasets while varying factors such as the amount of memory, number of CPU threads, storage device, GPU generation etc on servers that are a part of a large production cluster at Microsoft. We find that in many cases, DNN training time is dominated by data stall time: time spent waiting for data to be fetched and preprocessed. (2) We build a tool, DS-Analyzer to precisely measure data stalls using a differential technique, and perform predictive what-if analysis on data stalls. (3) Finally, based on the insights from our analysis, we design and implement three simple but effective techniques in a data-loading library, CoorDL, to mitigate data stalls. Our experiments on a range of DNN tasks, models, datasets, and hardware configs show that when PyTorch uses CoorDL instead of the state-of-the-art DALI data loading library, DNN training time is reduced significantly (by as much as 5x on a single server).

Supreeth Shastri, Melissa Wasserman, Vijay Chidambaram

In recent years, our society is being plagued by unprecedented levels of privacy and security breaches. To rein in this trend, the European Union, in 2018, introduced a comprehensive legislation called the General Data Protection Regulation (GDPR). In this paper, we review GDPR from a system design perspective, and identify how its regulations conflict with the design, architecture, and operation of modern systems. We illustrate these conflicts via the seven GDPR sins: storing data forever; reusing data indiscriminately; walled gardens and black markets; risk-agnostic data processing; hiding data breaches; making unexplainable decisions; treating security as a secondary goal. Our findings reveal a deep-rooted tussle between GDPR requirements and how modern systems have evolved. We believe that achieving compliance requires comprehensive, grounds up solutions, and anything short would amount to fixing a leaky faucet in a sinking ship.