This paper presents Tofu, a system that partitions very large DNN models across multiple GPU devices to reduce per-GPU memory footprint. Tofu is designed to partition a dataflow graph of fine-grained tensor operators used by platforms like MXNet and TensorFlow. In order to automatically partition each operator, we propose to describe the semantics of an operator in a simple language inspired by Halide. To optimally partition different operators in a dataflow graph, Tofu uses a recursive search algorithm that minimizes the total communication cost. Our experiments on an 8-GPU machine show that Tofu enables the training of very large CNN and RNN models. It also achieves 25% - 400% speedup over alternative approaches to train very large models.
more »
« less
SwapAdvisor: Pushing Deep Learning Beyond the GPU Memory Limit via Smart Swapping
It is known that deeper and wider neural networks can achieve better accuracy. But it is difficult to continue the trend to increase model size due to limited GPU memory. One promising solution is to support swapping between GPU and CPU memory. However, existing work on swapping only handle certain models and do not achieve satisfactory performance.
Deep learning computation is commonly expressed as a dataflow graph which can be analyzed to improve swapping. We propose SwapAdvisor, which performs joint optimization along 3 dimensions based on a given dataflow graph: operator scheduling, memory allocation, and swap decisions. SwapAdvisor explores the vast search space using a custom-designed genetic algorithm. Evaluations using a variety of large models show that SwapAdvisor can train models up to 12 times the GPU memory limit while achieving 53-99% of the throughput of a hypothetical baseline with infinite
GPU memory.
more »
« less
- Award ID(s):
- 1816717
- NSF-PAR ID:
- 10191573
- Date Published:
- Journal Name:
- International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS)
- Page Range / eLocation ID:
- 1341 to 1355
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Batching has a fundamental influence on the efficiency of deep neural network (DNN) execution. However, for dynamic DNNs, efficient batching is particularly challenging as the dataflow graph varies per input instance. As a result, state-of-the-art frameworks use heuristics that result in suboptimal batching decisions. Further, batching puts strict restrictions on memory adjacency and can lead to high data movement costs. In this paper, we provide an approach for batching dynamic DNNs based on finite state machines, which enables the automatic discovery of batching policies specialized for each DNN via reinforcement learning. Moreover, we find that memory planning that is aware of the batching policy can save significant data movement overheads, which is automated by a PQ tree-based algorithm we introduce. Experimental results show that our framework speeds up state-of-the-art frameworks by on average 1.15x, 1.39x, and 2.45x for chain-based, tree-based, and lattice-based DNNs across CPU and GPU. The framework is open-sourced at https://github.com/gulang2019/ED-Batch.git.more » « less
-
With the increase in the computation intensity of the chip, the mismatch between computation layer shapes and the available computation resource significantly limits the utilization of the chip. Driven by this observation, prior works discuss spatial accelerators or dataflow architecture to maximize the throughput. However, using spatial accelerators could potentially increase the execution latency. In this work, we first systematically investigate two execution models: (1) sequentially (temporally) launch one monolithic accelerator, and (2) spatially launch multiple accelerators. From the observations, we find that there is a latency throughput tradeoff between these two execution models, and combining these two strategies together can give us a more efficient latency throughput Pareto front. To achieve this, we propose spatial sequential architecture (SSR) and SSR design automation framework to explore both strategies together when deploying deep learning inference. We use the 7nm AMD Versal ACAP VCK190 board to implement SSR accelerators for four end-to-end transformer-based deep learning models. SSR achieves average throughput gains of 2.53x, 35.71x, and 14.20x under different batch sizes compared to the 8nm Nvidia GPU A10G, 16nm AMD FPGAs ZCU102, and U250. The average energy efficiency gains are 8.51x, 6.75x, and 21.22x, respectively. Compared with the sequential-only solution and spatial-only solution on VCK190, our spatial-sequential-hybrid solutions achieve higher throughput under the same latency requirement and lower latency under the same throughput requirement. We also use SSR analytical models to demonstrate how to use SSR to optimize solutions on other computing platforms, e.g., 14nm Intel Stratix 10 NX.more » « less
-
Recently, several task-parallel programming models have emerged to address the high synchronization and load imbalance issues as well as data movement overheads in modern shared memory architectures. OpenMP, the most commonly used shared memory parallel programming model, has added task execution support with dataflow dependencies. HPX and Regent are two more recent runtime systems that also support the dataflow execution model and extend it to distributed memory environments. We focus on parallelization of sparse matrix computations on shared memory architectures. We evaluate the OpenMP, HPX and Regent runtime systems in terms of performance and ease of implementation, and compare them against the traditional BSP model for two popular eigensolvers, Lanczos and LOBPCG. We give a general outline in regards to achieving parallelism using these runtime systems, and present a heuristic for tuning their performance to balance tasking overheads with the degree of parallelism that can be exposed. We then demonstrate their merits on two architectures, Intel Broadwell (a multicore processor) and AMD EPYC (a modern manycore processor). We observe that these frameworks achieve up to 13.7 × fewer cache misses over an efficient BSP implementation across L1, L2 and L3 cache layers. They also obtain up to 9.9 × improvement in execution time over the same BSP implementation.more » « less
-
We present a single-node, multi-GPU programmable graph processing library that allows programmers to easily extend single-GPU graph algorithms to achieve scalable performance on large graphs with billions of edges. Directly using the single-GPU implementations, our design only requires programmers to specify a few algorithm-dependent concerns, hiding most multi-GPU related implementation details. We analyze the theoretical and practical limits to scalability in the context of varying graph primitives and datasets. We describe several optimizations, such as direction optimizing traversal, and a just-enough memory allocation scheme, for better performance and smaller memory consumption. Compared to previous work, we achieve best-of-class performance across operations and datasets, including excellent strong and weak scalability on most primitives as we increase the number of GPUs in the system.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- https://doi.org/10.1145/3373376.3378530
