the success of Deep Learning (DL) algorithms in computer vision tasks have created an on-going demand of dedicated hardware architectures that could keep up withthe their required computation and memory complexities....
详细信息
ISBN:
(纸本)9781450371896
the success of Deep Learning (DL) algorithms in computer vision tasks have created an on-going demand of dedicated hardware architectures that could keep up withthe their required computation and memory complexities. this task is particularly challenging when embedded smart camera platforms have constrained resources such as power consumption, Processing Element (PE) and communication. this article describes a heterogeneous system embedding an FPGA and a GPU for executing CNN inference for computer vision applications. the built system addresses some challenges of embedded CNN such as task and data partitioning, and workload balancing. the selected heterogeneous platform embeds an Nvidia® Jetson TX2 for the CPU-GPU side and an Intel Altera® Cyclone10GX for the FPGA side interconnected by PCIe Gen2 with a MIPI-CSI camera for prototyping. this test environment will be used as a support for future work on a methodology for optimized model partitioning.
Huge energy consumption of large-scale cloud data centers damages environments with excessive carbon emission. More and more data center operators are seeking to reduce carbon footprint via various types of renewable ...
详细信息
Huge energy consumption of large-scale cloud data centers damages environments with excessive carbon emission. More and more data center operators are seeking to reduce carbon footprint via various types of renewable energy. However, the intermittent availability of renewable energy sources makes it quite challenging to cooperate with dynamically arriving workload. Meanwhile, the different natures (eg, price and carbon emission) of multiple energy sources also bring more challenges to achieve an optimal trade-off among carbon emission, power cost, and service level agreement (SLA). In this paper, we study the problem of reducing the long-term energy cost for geo-distributed cloud centers, where multiple sources of renewable energy are considered and SLA requirement and carbon budget are satisfied. To tackle the randomness of workload arrival, varying electricity price, and intermittent supply of renewable energy, we first formulate the cost minimization problem as a constraint stochastic optimization problem. Second, based on Lyapunov optimization technique, we propose an online control algorithm to solve it and provide the rigorous theory analysis to demonstrate its performance. By converting the long-term optimization problem to a mixed integer linear programming problem in each time slot, we analyze its inherent structure and propose an efficient algorithm to solve it based on Brenner's method. Our proposed algorithm makes online decisions rely only on the current system state and achieve [O(1V),O(V)] cost emission trade-off. Finally, the effectiveness of our algorithm is evaluated by extensive simulations based on real-world data traces.
A new emerging test and measurement standard called AXIe, AdvancedTCA eXtensions for Instrumentation, is expected to find wide acceptance within the Automatic Test Equipment community as it offers many key benefits. I...
详细信息
A new emerging test and measurement standard called AXIe, AdvancedTCA eXtensions for Instrumentation, is expected to find wide acceptance within the Automatic Test Equipment community as it offers many key benefits. It is expected that a large number of stimulus, measurement, signal conditioning, acquisition and processing modules will become available from a range of different suppliers. AXIe uses AdvancedTCA ® as its base standard, but then borrows from test and measurement industry standards such as PXI, IVI, and LXI, which were designed to facilitate cooperation and plug-and-play interoperability between instrument suppliers. this enables AXIe systems to easily integrate with other test and measurement equipment. AXIe's large board footprint, available power and efficient cooling to the module payload allows high density in a 19-inch rack space, enabling the development of high-performance instrumentation in a density unmatched by other instrumentation form factors. Channel synchronization between modules is flexible and provided by AXIe's dual triggering structures: a parallel trigger bus, and radially-distributed, time-matched point-to-point trigger lines. Inter-module communication is also provided with a local bus between adjacent modules allowing data transfer rates up to 10 Gbits/s in each direction, for example between front-end digitizer modules and DSP banks. the AXIe form factor provides the power and cooling necessary to embed high performance computing. A range of compute blades are available today in an AdvancedTCA ® form factor that provide low cost alternatives to the development of custom signal processing modules. the availability of both LAN and PCIe (PCI Express) fabrics allow the interconnection between modules, as well as high industry-standard highperformance data paths to external host computer systems. AXIe delivers a powerful environment for custom module development for specific and unique applications. As in the case of VXIbus and PX
In the "Grid" vision, demanding applications are dynamically assigned to supercomputers whose resources in terms of available CPU power and storage capacity ("grid resources") allow them to process...
详细信息
ISBN:
(纸本)9781424426256
In the "Grid" vision, demanding applications are dynamically assigned to supercomputers whose resources in terms of available CPU power and storage capacity ("grid resources") allow them to process the given jobs. However, distributedapplications may also be demanding in terms of the minimum bandwidth and maximum end-to-end delay ("network resources") available for the communication between its instances. this has lead to the development of several Network Resource Provisioning Systems (NRPSs), each offering certain possibilities to reserve bandwidth between network endpoints in a single administrative domain. To offer a greater variety of services to users and to facilitate cooperation among users of different domains, it seems natural to establish dedicated inter-domain links to extend the possibilities given by NRPSs to multiple domains. Unfortunately, it is not viable to place several domains under the control of any single NRPS to allow for inter-domain bandwidth reservations, since these systems generally require full network topology knowledge and administrative rights in each domain. Furthermore, each system was developed with a specific focus and therefore best fulfils the requirements of the users within certain domains. When considering possible approaches to enable multi-domain bandwidth reservations, it is important to have in mind the separation between "resource management", which lies within the autonomy of each domain and whose details are hidden to the outside, and the interoperability between the domains. this extended abstract will describe the design decisions taken to solve the exposed problems within the Phosphorus project.
暂无评论