Block propagation is a critical step in the consensus process, which determines the fork rate and transaction throughput of public blockchain systems. To accelerate block propagation, existing block relay protocols reduce the block size using transaction hashes, which requires the receiver to reconstruct the block based on the transactions in its mempool. Hence, their performance is highly affected by the number of transactions missed by mempools, especially in the P2P network with frequent arrival and departure of nodes. In this paper, we introduce Presync, a transaction synchronization protocol that can reduce the difference of transactions between the block and the mempool with controllable bandwidth overhead. It allows mining pool servers to synchronize the transactions in candidate blocks before the propagation of a valid block. High-bandwidth mode conducts a full synchronization of the candidate block using short hashes, and the Merkle root is utilized to match the valid block. Low-bandwidth mode provides a lightweight synchronization by identifying the unsynchronized transactions, so that the missing transactions can be detected with a low redundancy. We study the performance of Presync through the stochastic modeling and experimental evaluations. The results illustrate that high and low-bandwidth modes can respectively reduce the end-to-end delay of compact block by 60% and 78% with bandwidth usages 25KB and 63KB, in a network with 5 active pool servers and 2/3 online probability of full nodes.

Beamforming technique can effectively improve the spectrum utilization of multi-antenna systems, while the dirty-paper coding (DPC) technique can reduce inter-user interference. In this letter, we aim to maximize the weighted sum-rate under power constraint in a multiple-input-single-output (MISO) system with the DPC. However, the existing methods of beamforming optimization mainly rely on customized iterative algorithms, which have high computational complexity. To address this issue, by utilizing the deep learning technique and the uplink-downlink duality, and carefully exploring the optimal solution structure, we devise a beamforming neural network (BFNNet), which includes a deep neural network module and a signal processing module. Besides, we use the modulus of the channel coefficients as the input of deep neural network, which reduces the input size. Simulation results show that a well-trained BFNNet can achieve near-optimal solutions, while significantly reducing computational complexity