The state-of-the-art approaches employ approximate computing to reduce the energy consumption of DNN hardware. Approximate DNNs then require extensive retraining afterwards to recover from the accuracy loss caused by ...
详细信息
ISBN:
(纸本)9781728123509
The state-of-the-art approaches employ approximate computing to reduce the energy consumption of DNN hardware. Approximate DNNs then require extensive retraining afterwards to recover from the accuracy loss caused by the use of approximate operations. However, retraining of complex DNNs does not scale well. In this paper, we demonstrate that efficient approximations can be introduced into the computational path of DNN accelerators while retraining can completely be avoided. ALWANN provides highly optimized implementations of DNNs for custom low-power accelerators in which the number of computing units is lower than the number of DNN layers. First, a fully trained DNN (e.g., in TensorFlow) is converted to operate with 8-bit weights and 8-bit multipliers in convolutional layers. A suitable approximate multiplier is then selected for each computing element from a library of approximate multipliers in such a way that (i) one approximate multiplier serves several layers, and (ii) the overall classification error and energy consumption are minimized. The optimizations including the multiplier selection problem are solved by means of a multi-objective optimization NSGA-II algorithm. In order to completely avoid the computationally expensive retraining of DNN, which is usually employed to improve the classification accuracy, we propose a simple weight updating scheme that compensates the inaccuracy introduced by employing approximate multipliers. The proposed approach is evaluated for two architectures of DNN accelerators with approximate multipliers from the open-source "EvoApprox" library, while executing three versions of ResNet on CIFAR-10. We report that the proposed approach saves 30% of energy needed for multiplication in convolutional layers of ResNet-50 while the accuracy is degraded by only 0.6% (0.9% for the ResNet-14).
In the interconnected globe where service delivery is the success measure, cloud high availability (HA) is an indispensable area for enterprises. An HA-aware cloud system provides different approaches to handle the ou...
详细信息
In the interconnected globe where service delivery is the success measure, cloud high availability (HA) is an indispensable area for enterprises. An HA-aware cloud system provides different approaches to handle the outages. This includes geo-redundancy, failover schemes, and HA-aware placement solutions. However, using real-cloud platforms to model HA-aware approaches is hindered by the configuration settings. To this end, simulation tools, such as CloudSim, can be used to evaluate HA solutions and a cloud resiliency against failures. CloudSim allows implementing of scheduling policies, but it does not support HA properties. This paper provides availability-aware CloudSim extension (ACE). ACE extends CloudSim with a graphical and textual modeling to ensure simplicity and reusability of cloud scenarios. ACE has added HA-aware modeling (HA metrics and failure/redundancy/interdependency models) and HA-aware scheduling (HA-aware placements, failover, repair, and load balancing policies) into CloudSim. With ACE, the creation of cloud scenarios is facilitated, and multiple HA-aware deployment solutions can be evaluated under different stochastic and deterministic events. ACE can assess the impact of different redundancy/failure models, and other performance policies to extract HA-aware lessons. In this paper, ACE is assessed on a cloud application to evaluate different redundancy/failure models and provide availability analysis of the HA-aware placement solution.
Cloud computing is continuously growing as a business model for hosting information and communication technology applications. However, many concerns arise regarding the quality of service (QoS) offered by the cloud. ...
详细信息
Cloud computing is continuously growing as a business model for hosting information and communication technology applications. However, many concerns arise regarding the quality of service (QoS) offered by the cloud. One major challenge is the high availabil- ity (HA) of cloud-based applications. The key to achieving availability requirements is to develop an approach that is immune to cloud failures while minimizing the service level agreement (SLA) violations. To this end, this thesis addresses the HA of cloud-based applications from different per- spectives. First, the thesis proposes a component s HA-ware scheduler (CHASE) to man- age the deployments of carrier-grade cloud applications while maximizing their HA and satisfying the QoS requirements. Second, a Stochastic Petri Net (SPN) model is proposed to capture the stochastic characteristics of cloud services and quantify the expected avail- ability offered by an application deployment. The SPN model is then associated with an extensible policy-driven cloud scoring system that integrates other cloud challenges (i. e. green and cost concerns) with HA objectives. The proposed HA-aware solutions are ex- tended to include a live virtual machine migration model that provides a trade-off between the migration time and the downtime while maintaining HA objective. Furthermore, the thesis proposes a generic input template for cloud simulators, GITS, to facilitate the cre- ation of cloud scenarios while ensuring reusability, simplicity, and portability. Finally, an availability-aware CloudSim extension, ACE, is proposed. ACE extends CloudSim simu- lator with failure injection, computational paths, repair, failover, load balancing, and other availability-based modules.
In finite element analysis of pressure vessels undergoing elastoplastic deformation, low stiffness of the tangent modulus tensor will engender low stiffness in the tangent stiffness matrix, posing a risk of computatio...
详细信息
In finite element analysis of pressure vessels undergoing elastoplastic deformation, low stiffness of the tangent modulus tensor will engender low stiffness in the tangent stiffness matrix, posing a risk of computational difficulties such as poor convergence. The current investigation presents the explicit tangent modulus tensor in an elastoplastic model based on a Von Mises yield surface with isotropic work hardening, and the associated flow rule. The stiffness of the tangent modulus tensor is assessed by deriving explicit expressions for its minimum eigenvalue using both tensor diagonalization and Rayleigh quotient minimization. The derived expressions are validated computationally. Using the minimum eigenvalue, the stiffness is found to depend on the current path in stress space. The results of the current investigation suggest a way of following a stress path, which bypasses low stiffness, while attaining the prescribed load. [DOI: 10.1115/1.4004619]
暂无评论