Subscribe to the gold package and get unlimited access to Shamra Academy
Register a new userAnalysis of Large-Scale Multi-Tenant GPU Clusters for DNN Training Workloads
112
0
0.0
(
0
)
Added by
Myeongjae Jeon
Publication date
2019
fields
Informatics Engineering
and research's language is
English
Ask ChatGPT about the research
No Arabic abstract
With widespread advances in machine learning, a number of large enterprises are beginning to incorporate machine learning models across a number of products. These models are typically trained on shared, multi-tenant GPU clusters. Similar to existing cluster computing workloads, scheduling frameworks aim to provide features like high efficiency, resource isolation, fair sharing across users, etc. However Deep Neural Network (DNN) based workloads, predominantly trained on GPUs, differ in two significant ways from traditional big data analytics workloads. First, from a cluster utilization perspective, GPUs represent a monolithic resource that cannot be shared at a fine granularity across users. Second, from a workload perspective, deep learning frameworks require gang scheduling reducing the flexibility of scheduling and making the jobs themselves inelastic to failures at runtime. In this paper we present a detailed workload characterization of a two-month long trace from a multi-tenant GPU cluster in a large enterprise. By correlating scheduler logs with logs from individual jobs, we study three distinct issues that affect cluster utilization for DNN training workloads on multi-tenant clusters: (1) the effect of gang scheduling and locality constraints on queuing, (2) the effect of locality on GPU utilization, and (3) failures during training. Based on our experience running a large-scale operation, we provide design guidelines pertaining to next-generation cluster schedulers for DNN training workloads.
rate research
Read More
Systems for processing big data---e.g., Hadoop, Spark, and massively parallel databases---need to run workloads on behalf of multiple tenants simultaneously. The abundant disk-based storage in these systems is usually complemented by a smaller, but much faster, {em cache}. Cache is a precious resource: Tenants who get to use cache can see two orders of magnitude performance improvement. Cache is also a limited and hence shared resource: Unlike a resource like a CPU core which can be used by only one tenant at a time, a cached data item can be accessed by multiple tenants at the same time. Cache, therefore, has to be shared by a multi-tenancy-aware policy across tenants, each having a unique set of priorities and workload characteristics. In this paper, we develop cache allocation strategies that speed up the overall workload while being {em fair} to each tenant. We build a novel fairness model targeted at the shared resource setting that incorporates not only the more standard concepts of Pareto-efficiency and sharing incentive, but also define envy freeness via the notion of {em core} from cooperative game theory. Our cache management platform, ROBUS, uses randomization over small time batches, and we develop a proportionally fair allocation mechanism that satisfies the core property in expectation. We show that this algorithm and related fair algorithms can be approximated to arbitrary precision in polynomial time. We evaluate these algorithms on a ROBUS prototype implemented on Spark with RDD store used as cache. Our evaluation on a synthetically generated industry-standard workload shows that our algorithms provide a speedup close to performance optimal algorithms while guaranteeing fairness across tenants.
Modern GPU datacenters are critical for delivering Deep Learning (DL) models and services in both the research community and industry. When operating a datacenter, optimization of resource scheduling and management can bring significant financial benefits. Achieving this goal requires a deep understanding of the job features and user behaviors. We present a comprehensive study about the characteristics of DL jobs and resource management. First, we perform a large-scale analysis of real-world job traces from SenseTime. We uncover some interesting conclusions from the perspectives of clusters, jobs and users, which can facilitate the cluster system designs. Second, we introduce a general-purpose framework, which manages resources based on historical data. As case studies, we design: a Quasi-Shortest-Service-First scheduling service, which can minimize the cluster-wide average job completion time by up to 6.5x; and a Cluster Energy Saving service, which improves overall cluster utilization by up to 13%.
We present a Gauss-Newton-Krylov solver for large deformation diffeomorphic image registration. We extend the publicly available CLAIRE library to multi-node multi-graphics processing unit (GPUs) systems and introduce novel algorithmic modifications that significantly improve performance. Our contributions comprise ($i$) a new preconditioner for the reduced-space Gauss-Newton Hessian system, ($ii$) a highly-optimized multi-node multi-GPU implementation exploiting device direct communication for the main computational kernels (interpolation, high-order finite difference operators and Fast-Fourier-Transform), and ($iii$) a comparison with state-of-the-art CPU and GPU implementations. We solve a $256^3$-resolution image registration problem in five seconds on a single NVIDIA Tesla V100, with a performance speedup of 70% compared to the state-of-the-art. In our largest run, we register $2048^3$ resolution images (25 B unknowns; approximately 152$times$ larger than the largest problem solved in state-of-the-art GPU implementations) on 64 nodes with 256 GPUs on TACCs Longhorn system.
Efficient GPU resource scheduling is essential to maximize resource utilization and save training costs for the increasing amount of deep learning workloads in shared GPU clusters. Existing GPU schedulers largely rely on static policies to leverage the performance characteristics of deep learning jobs. However, they can hardly reach optimal efficiency due to the lack of elasticity. To address the problem, we propose ONES, an ONline Evolutionary Scheduler for elastic batch size orchestration. ONES automatically manages the elasticity of each job based on the training batch size, so as to maximize GPU utilization and improve scheduling efficiency. It determines the batch size for each job through an online evolutionary search that can continuously optimize the scheduling decisions. We evaluate the effectiveness of ONES with 64 GPUs on TACCs Longhorn supercomputers. The results show that ONES can outperform the prior deep learning schedulers with a significantly shorter average job completion time.
Large language models have led to state-of-the-art accuracies across a range of tasks. However, training these models efficiently is challenging for two reasons: a) GPU memory capacity is limited, making it impossible to fit large models on even a multi-GPU server, and b) the number of compute operations required to train these models can result in unrealistically long training times. Consequently, new methods of model parallelism such as tensor and pipeline parallelism have been proposed. Unfortunately, naive usage of these methods leads to fundamental scaling issues at thousands of GPUs, e.g., due to expensive cross-node communication or devices spending significant time waiting on other devices to make progress. In this paper, we show how different types of parallelism methods (tensor, pipeline, and data parallelism) can be composed to scale to thousands of GPUs and models with trillions of parameters. We survey techniques for pipeline parallelism and propose a novel interleaved pipeline parallelism schedule that can improve throughput by 10+% with memory footprint comparable to existing approaches. We quantitatively study the trade-offs between tensor, pipeline, and data parallelism, and provide intuition as to how to configure distributed training of a large model. Our approach allows us to perform training iterations on a model with 1 trillion parameters at 502 petaFLOP/s on 3072 GPUs with achieved per-GPU throughput of 52% of theoretical peak. Our code is open sourced at https://github.com/nvidia/megatron-lm.
Log in to be able to interact and post comments
comments
Fetching comments
Sign in to be able to follow your search criteria
