arXiv subjects

Deepak Narayanan

Publications and source records attributed to Deepak Narayanan.

29 records · Page 2Linked to original sources

arXiv2021

Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-LM

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.

cs.CL↗

arXiv2021

Memory-Efficient Pipeline-Parallel DNN Training

Many state-of-the-art ML results have been obtained by scaling up the number of parameters in existing models. However, parameters and activations for such large models often do not fit in the memory of a single accelerator device; this means that it is necessary to distribute training of large models over multiple accelerators. In this work, we propose PipeDream-2BW, a system that supports memory-efficient pipeline parallelism. PipeDream-2BW uses a novel pipelining and weight gradient coalescing strategy, combined with the double buffering of weights, to ensure high throughput, low memory footprint, and weight update semantics similar to data parallelism. In addition, PipeDream-2BW automatically partitions the model over the available hardware resources, while respecting hardware constraints such as memory capacities of accelerators and interconnect topologies. PipeDream-2BW can accelerate the training of large GPT and BERT language models by up to 20$\times$ with similar final model accuracy.

cs.LG↗

arXiv2021

Allocation of Fungible Resources via a Fast, Scalable Price Discovery Method

We consider the problem of assigning or allocating resources to a set of jobs. We consider the case when the resources are fungible, that is, the job can be done with any mix of the resources, but with different efficiencies. In our formulation we maximize a total utility subject to a given limit on the resource usage, which is a convex optimization problem and so is tractable. In this paper we develop a custom, parallelizable algorithm for solving the resource allocation problem that scales to large problems, with millions of jobs. Our algorithm is based on the dual problem, in which the dual variables associated with the resource usage limit can be interpreted as resource prices. Our method updates the resource prices in each iteration, ultimately discovering the optimal resource prices, from which an optimal allocation is obtained. We provide an open-source implementation of our method, which can solve problems with millions of jobs in a few seconds on CPU, and under a second on a GPU; our software can solve smaller problems in milliseconds. On large problems, our implementation is up to three orders of magnitude faster than a commerical solver for convex optimization.

math.OC↗

arXiv2021

Don't Give Up on Large Optimization Problems; POP Them!

Resource allocation problems in many computer systems can be formulated as mathematical optimization problems. However, finding exact solutions to these problems using off-the-shelf solvers in an online setting is often intractable for "hyper-scale" system sizes with tight SLAs, leading system designers to rely on cheap, heuristic algorithms. In this work, we explore an alternative approach that reuses the original optimization problem formulation. By splitting the original problem into smaller, more tractable problems for subsets of the system and then coalescing resulting sub-allocations into a global solution, we achieve empirically quasi-optimal (within 1.5%) performance for multiple domains with several orders-of-magnitude improvement in runtime. Deciding how to split a large problem into smaller sub-problems, and how to coalesce split allocations into a unified allocation, needs to be performed carefully in a domain-aware way. We show common principles for splitting problems effectively across a variety of tasks, including cluster scheduling, traffic engineering, and load balancing.

cs.DC↗

arXiv2020

Heterogeneity-Aware Cluster Scheduling Policies for Deep Learning Workloads

Specialized accelerators such as GPUs, TPUs, FPGAs, and custom ASICs have been increasingly deployed to train deep learning models. These accelerators exhibit heterogeneous performance behavior across model architectures. Existing schedulers for clusters of accelerators, which are used to arbitrate these expensive training resources across many users, have shown how to optimize for various multi-job, multi-user objectives, like fairness and makespan. Unfortunately, existing schedulers largely do not consider performance heterogeneity. In this paper, we propose Gavel, a heterogeneity-aware scheduler that systematically generalizes a wide range of existing scheduling policies. Gavel expresses these policies as optimization problems, making it easy to optimize for objectives in a heterogeneity-aware way, while also being cognizant of performance optimizations like space sharing. Gavel then uses a round-based scheduling mechanism to ensure jobs receive their ideal allocation given the target scheduling policy. Gavel's heterogeneity-aware policies allow a heterogeneous cluster to sustain higher input load, and improve end objectives such as average job completion time and makespan by up to 3.5x compared to heterogeneity-agnostic policies.

cs.DC↗

arXiv2020

Willump: A Statistically-Aware End-to-end Optimizer for Machine Learning Inference

Systems for ML inference are widely deployed today, but they typically optimize ML inference workloads using techniques designed for conventional data serving workloads and miss critical opportunities to leverage the statistical nature of ML. In this paper, we present Willump, an optimizer for ML inference that introduces two statistically-motivated optimizations targeting ML applications whose performance bottleneck is feature computation. First, Willump automatically cascades feature computation for classification queries: Willump classifies most data inputs using only high-value, low-cost features selected through empirical observations of ML model performance, improving query performance by up to 5x without statistically significant accuracy loss. Second, Willump accurately approximates ML top-K queries, discarding low-scoring inputs with an automatically constructed approximate model and then ranking the remainder with a more powerful model, improving query performance by up to 10x with minimal accuracy loss. Willump automatically tunes these optimizations' parameters to maximize query performance while meeting an accuracy target. Moreover, Willump complements these statistical optimizations with compiler optimizations to automatically generate fast inference code for ML applications. We show that Willump improves the end-to-end performance of real-world ML inference pipelines curated from major data science competitions by up to 16x without statistically significant loss of accuracy.

cs.DB↗

arXiv2020

MLPerf Training Benchmark

Machine learning (ML) needs industry-standard performance benchmarks to support design and competitive evaluation of the many emerging software and hardware solutions for ML. But ML training presents three unique benchmarking challenges absent from other domains: optimizations that improve training throughput can increase the time to solution, training is stochastic and time to solution exhibits high variance, and software and hardware systems are so diverse that fair benchmarking with the same binary, code, and even hyperparameters is difficult. We therefore present MLPerf, an ML benchmark that overcomes these challenges. Our analysis quantitatively evaluates MLPerf's efficacy at driving performance and scalability improvements across two rounds of results from multiple vendors.

cs.LG↗

arXiv2019

Analysis of DAWNBench, a Time-to-Accuracy Machine Learning Performance Benchmark

Researchers have proposed hardware, software, and algorithmic optimizations to improve the computational performance of deep learning. While some of these optimizations perform the same operations faster (e.g., increasing GPU clock speed), many others modify the semantics of the training procedure (e.g., reduced precision), and can impact the final model's accuracy on unseen data. Due to a lack of standard evaluation criteria that considers these trade-offs, it is difficult to directly compare these optimizations. To address this problem, we recently introduced DAWNBench, a benchmark competition focused on end-to-end training time to achieve near-state-of-the-art accuracy on an unseen dataset---a combined metric called time-to-accuracy (TTA). In this work, we analyze the entries from DAWNBench, which received optimized submissions from multiple industrial groups, to investigate the behavior of TTA as a metric as well as trends in the best-performing entries. We show that TTA has a low coefficient of variation and that models optimized for TTA generalize nearly as well as those trained using standard methods. Additionally, even though DAWNBench entries were able to train ImageNet models in under 3 minutes, we find they still underutilize hardware capabilities such as Tensor Cores. Furthermore, we find that distributed entries can spend more than half of their time on communication. We show similar findings with entries to the MLPERF v0.5 benchmark.

cs.LG↗

arXiv2018

PipeDream: Fast and Efficient Pipeline Parallel DNN Training

PipeDream is a Deep Neural Network(DNN) training system for GPUs that parallelizes computation by pipelining execution across multiple machines. Its pipeline parallel computing model avoids the slowdowns faced by data-parallel training when large models and/or limited network bandwidth induce high communication-to-computation ratios. PipeDream reduces communication by up to 95% for large DNNs relative to data-parallel training, and allows perfect overlap of communication and computation. PipeDream keeps all available GPUs productive by systematically partitioning DNN layers among them to balance work and minimize communication, versions model parameters for backward pass correctness, and schedules the forward and backward passes of different inputs in round-robin fashion to optimize "time to target accuracy". Experiments with five different DNNs on two different clusters show that PipeDream is up to 5x faster in time-to-accuracy compared to data-parallel training.

cs.DC↗

arXiv2017

Weld: Rethinking the Interface Between Data-Intensive Applications

Data analytics applications combine multiple functions from different libraries and frameworks. Even when each function is optimized in isolation, the performance of the combined application can be an order of magnitude below hardware limits due to extensive data movement across these functions. To address this problem, we propose Weld, a new interface between data-intensive libraries that can optimize across disjoint libraries and functions. Weld exposes a lazily-evaluated API where diverse functions can submit their computations in a simple but general intermediate representation that captures their data-parallel structure. It then optimizes data movement across these functions and emits efficient code for diverse hardware. Weld can be integrated into existing frameworks such as Spark, TensorFlow, Pandas and NumPy without changing their user-facing APIs. We demonstrate that Weld can speed up applications using these frameworks by up to 29x.

cs.DC↗

arXiv2017

MacroBase: Prioritizing Attention in Fast Data

As data volumes continue to rise, manual inspection is becoming increasingly untenable. In response, we present MacroBase, a data analytics engine that prioritizes end-user attention in high-volume fast data streams. MacroBase enables efficient, accurate, and modular analyses that highlight and aggregate important and unusual behavior, acting as a search engine for fast data. MacroBase is able to deliver order-of-magnitude speedups over alternatives by optimizing the combination of explanation and classification tasks and by leveraging a new reservoir sampler and heavy-hitters sketch specialized for fast data streams. As a result, MacroBase delivers accurate results at speeds of up to 2M events per second per query on a single core. The system has delivered meaningful results in production, including at a telematics company monitoring hundreds of thousands of vehicles.

cs.DB↗