Recent advances in distributed machine learning show theoretically and empirically that, for many models, provided that workers will eventually participate in the synchronizations, i) the training still converges, eve...
详细信息
Recent advances in distributed machine learning show theoretically and empirically that, for many models, provided that workers will eventually participate in the synchronizations, i) the training still converges, even if only p workers take part in each round of synchronization, and ii) a larger p generally leads to a faster rate of convergence. These findings shed light on eliminating the bottleneck effects of parameter synchronization in large-scale data-parallel distributed training and have motivated several optimization designs. In this paper, we focus on optimizing the parameter synchronization for peer-to-peer distributed learning, where workers broadcast or multicast their updated parameters to others for synchronization, and propose SELMCAST, a suite of expressive and efficient multicast receiver selection algorithms, to achieve the goal. Compared with the state-of-the-art (SOTA) design, which randomly selects exactly p receivers for each worker's multicast in a bandwidth-agnostic way, SELMCAST chooses receivers based on the global view of their available bandwidth and loads, yielding two advantages, i.e., accelerated parameter synchronization for higher utilization of computing resources and enlarged average p values for faster convergence. Comprehensive evaluations show that SELMCAST is efficient for both peer-to-peer Bulk Synchronous Parallel (BSP) and Stale Synchronous Parallel (SSP) distributed training, outperforming the SOTA solution significantly.
Recent advances in distributed machine learning show theoretically and empirically that, for many models, provided workers would participate in the synchronizations eventually, 8) the training still converges, even if...
详细信息
ISBN:
(数字)9781538683477
ISBN:
(纸本)9781538683477
Recent advances in distributed machine learning show theoretically and empirically that, for many models, provided workers would participate in the synchronizations eventually, 8) the training still converges, even if only ? workers take part in each round of synchronization, and 88) a larger ? generally leads to a faster rate of convergence. These findings shed light on eliminating the bottleneck effects of parameter synchronization in large-scale data-parallel distributed training, having motivated several optimization designs. In this paper, we focus on optimizing the parameter synchronization for peer-to-peer distributed learning, in which workers generally broadcast or multicast their updated parameters to others for synchronization, and propose SELMCAST, an expressive and Pareto-optimal multicast receiver selection algorithm, to achieve the goal. Compared with the state-of-the-art design that randomly selects exactly ? receivers for each worker's multicast in a bandwidth-agnostic way, SELMCAST chooses receivers based on the global view of their available bandwidth and loads, yielding two advantages. Firstly, it could optimize the bottleneck sending rate, thus cutting down the time cost of parameter synchronization. Secondly, when more than ? receivers are with sufficient bandwidth, they would be selected as many as possible, bringing benefits to the convergence of training. Extensive evaluations show that SELMCAST is efficient and always achieves near-optimal performance.
The discontinuity of tangent, curvature, and curvature derivative of the 5-axis toolpath consisting of line and arc segments defined by the G01/G02/G03 commands limits the performance of hybrid robot in high-efficienc...
详细信息
The discontinuity of tangent, curvature, and curvature derivative of the 5-axis toolpath consisting of line and arc segments defined by the G01/G02/G03 commands limits the performance of hybrid robot in high-efficiency and high-precision milling machining. To guarantee the high-order continuous and smooth machining, a C-3 continuous corner smoothing 5-axis toolpath containing spatial arc segments is semi-analytically generated. Firstly, the control points of the smoothing quintic B-spline for line and arc segments are constructed analytically based on the G(3) continuity conditions. Within the constraints of the established explicit smoothing error expressions, the G(3) continuous smoothing for three kinds of tooltip positions, namely, spatial line-line pairs, arc-arc pairs, and line-arc pairs, is investigated. Meanwhile, the smoothing error model of tool orientation is analytically established. Then, the parameter synchronization of the tool orientation at the junction following the tooltip position is achieved by replacing the remaining segments with seven B-spline curves as synchronization curves. Finally, simulations and experiments with the typical 5-axis toolpath containing spatial arc segments are carried out to verify the validity of the proposed method in ensuring C-3 continuous motion in the robot operation space and good following performance in the joint space.
A multibeam concept under full frequency reuse is a major prerequisite for high throughput satellite systems. The resulting interference problems might be tackled by appropriately designed precoding or multiuser detec...
详细信息
A multibeam concept under full frequency reuse is a major prerequisite for high throughput satellite systems. The resulting interference problems might be tackled by appropriately designed precoding or multiuser detection schemes. However, before such powerful techniques are applicable, the most important transmission parameters have to be recovered successfully. In the context of this paper, a feedback structure for joint control of carrier frequency and symbol timing is investigated for a multibeam scenario, which has been developed in some previous work by the authors for a single user link. It is to be noticed that we describe a recovery method for a multibeam satellite network with suitably selected interference mitigation techniques;as a consequence, all cochannels are assumed to be aligned in frequency and time to the reference beam;other signal models are out of scope and not addressed in this contribution.
In this paper, a discrete iterative model of feedback-type DC/DC converter is constructed using Laplace transform. The advantage of this model is proved comparing with state space average model. Self-adaptive control ...
详细信息
ISBN:
(纸本)9783642238864
In this paper, a discrete iterative model of feedback-type DC/DC converter is constructed using Laplace transform. The advantage of this model is proved comparing with state space average model. Self-adaptive control and parameter synchronization is applied to the chaos control of DC/DC converter. Simulation results show the effectiveness of this mathematical model and control method.
In a recently published paper, a novel approach for blind feedback synchronization of the symbol timing has been proposed. In order to avoid any jitter floor irrespective of the selected modulation scheme, a filter wi...
详细信息
ISBN:
(纸本)9781509016822
In a recently published paper, a novel approach for blind feedback synchronization of the symbol timing has been proposed. In order to avoid any jitter floor irrespective of the selected modulation scheme, a filter with very particular properties has been developed and placed in parallel to the receiver matched filter. Nevertheless, the jitter performance of this method degrades in case the carrier frequency offset does not vanish. Therefore, joint synchronization of symbol timing and carrier frequency will be studied in the current paper, based on the dual-filter framework developed in some previous works. To this end, we introduce a very simple frequency error detector, which is analyzed in terms of open-loop characteristic (S-curve) and jitter variance, forming this way the ground for joint recovery.
The linear-format path is widely adopted to approximate the continuous contour in robot controllers. The tangential discontinuity of the linear paths usually causes the discontinuity of the joint velocity. To comply w...
详细信息
The linear-format path is widely adopted to approximate the continuous contour in robot controllers. The tangential discontinuity of the linear paths usually causes the discontinuity of the joint velocity. To comply with the joint kinematics limits, the robots have to stop at each corner point, resulting in a great loss of efficiency. To achieve a smooth motion, this paper presents an analytical decoupled C3 continuous local path smoothing method for industrial robots. The tool position path is smoothed in the reference frame while the tool orientation is smoothed in the rotation parametric space based on the exponential coordinates of rotations. The quintic Bsplines are inserted at the corners of the linear segments to achieve the G3 continuity of the tool position path and tool orientation path. The orientation smoothing error is constrained analytically. By reparameterization of the remaining linear segments using specially constructed B-splines, the C3 continuity of the tool position path and tool orientation path is achieved. Then, the synchronization of the tool orientation path and tool position path can be guaranteed by sharing the same curve parameter. Besides, to improve the smoothness of the angular motion on the remaining linear segments during parameter synchronization, the transition lengths of the inserted B-splines are optimized. The proposed local smoothing method guarantees that the generated smooth orientation path in the rotation space is invariant with the selection of the reference frame and the orientation of the tool frame, and ensures the jerk-continuous motion with a smoother angular motion on the remaining linear segments, which could improve the motion smoothness of the tool path and tracking accuracy. The effectiveness of the proposed method is validated through simulation and experiments.
暂无评论