An official website of the United States government
With the success of deep neural networks (DNN), many recent works have been focusing on developing hardware accelerator for power and resource-limited embedded system via model compression techniques, such as quantization, pruning, low-rank approximation, etc. However, almost all existing DNN structure is fixed after deployment, which lacks runtime adaptive DNN structure to adapt to its dynamic hardware resource, power budget, throughput requirement, as well as dynamic workload. Correspondingly, there is no runtime adaptive hardware platform to support dynamic DNN structure. To address this problem, we first propose a dynamic channel-adaptive deep neural network (CA-DNN) which can adjust the involved convolution channel (i.e. model size, computing load) at run-time (i.e. at inference stage without retraining) to dynamically trade off between power, speed, computing load and accuracy. Further, we utilize knowledge distillation method to optimize the model and quantize the model to 8-bits and 16-bits, respectively, for hardware friendly mapping. We test the proposed model on CIFAR-10 and ImageNet dataset by using ResNet. Comparing with the same model size of individual model, our CA-DNN achieves better accuracy. Moreover, as far as we know, we are the first to propose a Processing-in-Memory accelerator for such adaptive neural networks structure based on Spin Orbit Torque Magnetic Random Access Memory(SOT-MRAM) computational adaptive sub-arrays. Then, we comprehensively analyze the trade-off of the model with different channel-width between the accuracy and the hardware parameters, eg., energy, memory, and area overhead.
more »
« less
Sample Transfer Optimization with Adaptive Deep Neural Network
Application-layer transfer configurations play a crucial role in achieving desirable performance in high-speed networks. However, finding the optimal configuration for a given transfer task is a difficult problem as it depends on various factors including dataset characteristics, network settings, and background traffic. The state-of-the-art transfer tuning solutions rely on real-time sample transfers to evaluate various configurations and estimate the optimal one. However, existing approaches to run sample transfers incur high delay and measurement errors, thus significantly limit the efficiency of the transfer tuning algorithms. In this paper, we introduce adaptive feed forward deep neural network (DNN) to minimize the error rate of sample transfers without increasing their execution time. We ran 115K file transfers in four different high-speed networks and used their logs to train an adaptive DNN that can quickly and accurately predict the throughput of sample transfers by analyzing instantaneous throughput values. The results gathered in various networks with rich set of transfer configurations indicate that the proposed model reduces error rate by up to 50% compared to the state-of-the-art solutions while keeping the execution time low. We also show that one can further reduce delay or error rate by tuning hyperparameters of the model to meet specific needs of user or application. Finally, transfer learning analysis reveals that the model developed in one network would yield accurate results in other networks with similar transfer convergence characteristics, alleviating the needs to run an extensive data collection and model derivation efforts for each network.
more »
« less
- Award ID(s):
- 1850353
- PAR ID:
- 10129011
- Date Published:
- Journal Name:
- IEEE/ACM Innovating the Network for Data-Intensive Science (INDIS)
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Efficiently transferring data over long-distance, high-speed networks requires optimal utilization of available network bandwidth. One effective method to achieve this is through the use of parallel TCP streams. This approach allows applications to leverage network parallelism, thereby enhancing transfer throughput. However, determining the ideal number of parallel TCP streams can be challenging due to non-deterministic background traffic sharing the network, as well as non-stationary and partially observable network signals. We present a novel learning-based approach that utilizes deep reinforcement learning (DRL) to determine the optimal number of parallel TCP streams. Our DRL-based algorithm is designed to intelligently utilize available network bandwidth while adapting to different network conditions. Unlike rule-based heuristics, which lack generalization in unknown network scenarios, our DRL-based solution can dynamically adjust the parallel TCP stream numbers to optimize network bandwidth utilization without causing network congestion and ensuring fairness among competing transfers. We conducted extensive experiments to evaluate our DRL-based algorithm’s performance and compared it with several state-of-the-art online optimization algorithms. The results demonstrate that our algorithm can identify nearly optimal solutions 40% faster while achieving up to 15% higher throughput. Furthermore, we show that our solution can prevent network congestion and distribute the available network resources fairly among competing transfers, unlike a discriminatory algorithm.more » « less
-
null (Ed.)With the success of Deep Neural Networks (DNN), many recent works have been focusing on developing hardware accelerator for power and resource-limited system via model compression techniques, such as quantization, pruning, low-rank approximation and etc. However, almost all existing compressed DNNs are fixed after deployment, which lacks run-time adaptive structure to adapt to its dynamic hardware resource allocation, power budget, throughput requirement, as well as dynamic workload. As the countermeasure, to construct a novel run-time dynamic DNN structure, we propose a novel DNN sub-network sampling method via non-uniform channel selection for subnets generation. Thus, user can trade off between power, speed, computing load and accuracy on-the-fly after the deployment, depending on the dynamic requirements or specifications of the given system. We verify the proposed model on both CIFAR-10 and ImageNet dataset using ResNets, which outperforms the same sub-nets trained individually and other related works. It shows that, our method can achieve latency trade-off among 13.4, 24.6, 41.3, 62.1(ms) and 30.5, 38.7, 51, 65.4(ms) for GPU with 128 batch-size and CPU respectively on ImageNet using ResNet18.more » « less
-
Deep neural network (DNN) accelerators as an example of domain-specific architecture have demonstrated great success in DNN inference. However, the architecture acceleration for equally important DNN training has not yet been fully studied. With data forward, error backward and gradient calculation, DNN training is a more complicated process with higher computation and communication intensity. Because the recent research demonstrates a diminishing specialization return, namely, “accelerator wall”, we believe that a promising approach is to explore coarse-grained parallelism among multiple performance-bounded accelerators to support DNN training. Distributing computations on multiple heterogeneous accelerators to achieve high throughput and balanced execution, however, remaining challenging. We present ACCPAR, a principled and systematic method of determining the tensor partition among heterogeneous accelerator arrays. Compared to prior empirical or unsystematic methods, ACCPAR considers the complete tensor partition space and can reveal previously unknown new parallelism configurations. ACCPAR optimizes the performance based on a cost model that takes into account both computation and communication costs of a heterogeneous execution environment. Hence, our method can avoid the drawbacks of existing approaches that use communication as a proxy of the performance. The enhanced flexibility of tensor partitioning in ACCPAR allows the flexible ratio of computations to be distributed among accelerators with different performances. The proposed search algorithm is also applicable to the emerging multi-path patterns in modern DNNs such as ResNet. We simulate ACCPAR on a heterogeneous accelerator array composed of both TPU-v2 and TPU-v3 accelerators for the training of large-scale DNN models such as Alexnet, Vgg series and Resnet series. The average performance improvements of the state-of-the-art “one weird trick” (OWT) and HYPAR, and ACCPAR, normalized to the baseline data parallelism scheme where each accelerator replicates the model and processes different input data in parallel, are 2.98×, 3.78×, and 6.30×, respectively.more » « less
-
null (Ed.)This paper presents CodedBulk, a system for high-throughput inter-datacenter bulk transfers. At its core, CodedBulk uses network coding, a technique from the coding theory community, that guarantees optimal throughput for individual bulk transfers. Prior attempts to using network coding in wired networks have faced several pragmatic and fundamental barriers. CodedBulk resolves these barriers by exploiting the unique properties of inter-datacenter networks, and by using a custom-designed hop-by-hop flow control mechanism that enables efficient realization of network coding atop existing transport protocols. An end-to end CodedBulk implementation running on a geo-distributed inter-datacenter network improves bulk transfer throughput by 1.2−2.5× compared to state-of-the-art mechanisms that do not use network coding.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- https://doi.org/10.1109/INDIS49552.2019.00013
