We present PHOTONS (Pose-Free Human-Centric Photo-Realistic Real-Time Novel View Synthesis from Sparse Views), a real-time framework for novel view synthesis without requiring camera calibration. Our method reconstructs consistent 3D Gaussian point clouds and synthesizes 2K photo-realistic novel views from arbitrary numbers (>=2) of freely placed cameras. PHOTONS faithfully renders dynamic human bodies amid complex backgrounds, including interactive object manipulation and fine-grained details (e.g., hair strands), while maintaining 25 FPS throughput on commodity GPU like NVIDIA RTX 4090. By combining pose-free spatial point cloud reconstruction with Gaussian parameter estimation, our method demonstrates strong resilience to occlusions and camera perturbations. Additionally, we develop a 3D stereo system that drastically reduces setup complexity compared to existing solutions. Experiments on public and custom datasets show that PHOTONS outperforms state-of-the-art methods in both efficiency and visual quality.

To ensure the data consistency, there is usually only one master service node in the cloud storage system that provides external services at the same time. Based on cost and efficiency considerations, users also hope to fully utilize the resources of each service node in resource-intensive execution environment. However, the master node selection approach in traditional distributed coordination services is imbalanced in weight. In this paper, we have proposed an equal opportunity approach based on VID for selecting master nodes towards resource-intensive cloud computing environment. Because the VID has a strict full order relationship, we could select the master node in one round, reducing the unavailable time of the cloud storage system during the process of reaching a consensus agreement. Through our proposed approach, we have solved the problem that some nodes frequently become the master, resulting in excessive machine load and device losses, while some nodes never have the opportunity to become the master, resulting in waste of idle resources.

Due to the characteristics of "erase-before-write" and "page write, block erase", SSDs will perform out-of-place updates. Out-of-place updates will produce a large number of fragmented pages, so the garbage collection process is required. In order to ensure the smooth progress of garbage collection, OP space needs to be set. If the OP space is set too large, the storage resources will be wasted. If the OP space is set too small, the write amplification and block erasure count will be increased, and the lifetime of SSDs will be reduced. To improve the overall utilization of SSDs, an innovative bridge layer access control method is designed. Firstly, the capacity mapping module provides real mapping and virtual mapping to host and operating system respectively, so that smaller OP space can be set and the space utilization of SSDs can be improved. Then, random writes are converted into sequential writes through the write conversion module, so as to reduce the number of block erasures and improve the lifetime of SSDs. Finally, the effectiveness of the proposed bridging layer is proved by experiments.

To ensure the security of data transmission on the Internet, users usually encrypt the data during the process of sending and receiving data. For example, the commonly used HTTPS protocol verifies the identity of servers through TLS certificates and encrypts communication between browsers and servers. However, the encrypted part of the unstructured data is still re-encrypted by TLS during HTTPS transmission, wasting computing and energy resources. In this paper, we propose an energy efficient network protocol, namely MRP, that avoids multiple encryption. MRP could carry multiple types of application layer protocols, while freely configure the location and encryption approaches of the data that needs to be encrypted. Based on our proposed, users could freely segment application layer data, achieve on-demand encryption of data, reduce encryption costs without compromising user security requirements, avoid redundant double encryption and save energy.

System data availability is a key indicator to measure the ability of a cloud storage system to provide continuous service. Usually, users could only get the data availability of a cloud storage system through measurement, meaning that they could not know whether the system works normally unless they have read and written data on it. To improve data availability, the common method in the industry is to use higher performance physical devices or design more complex organizational architectures. However, no matter how the service side is designed, if the client side does not have the ability to switch service points, the whole cloud storage system is not highly available. In this paper, we propose a data availability modeling approach towards cloud storage systems based on client perspective. Initially, we discuss traditional availability computing models and summarize the necessary conditions for a cloud storage system to realize high availability. Furthermore, combined with specific scenarios, we model two different modes of switching service nodes on the client side. Finally, we quantitatively analyze the optimal ratio of the detection requests and the read-write buffer on the premise of ensuring the storage networks performance.

Rapidly growing data volumes stimulate the migration of block storage from on-premise to hybrid cloud. Hybrid cloud block storage is composed of both on-premise and remote cloud storage. Unfortunately, the heterogeneity of on-premise clients (e.g., laptops, mobile devices, etc.) with various operating systems encumber the hybrid cloud architecture design. What's worse, the demanding fault-tolerance and latency requirements pose challenges to the block storage system. Under the premise of strong consistency and availability, there are two mainstream models of hybrid storage systems: the low latency but low throughput “active-standby” model where only one node can provide reads and writes, and the high throughput but high latency “active-active” model where all nodes can provide reads and writes. To address the dilemma between throughput and latency, we design a novel storage model called the “logical-active” model, which partitions the input data into shards in terms of the block address ranges and selects one owner node called “logical-active” for each shard. A unique “logical-active” node manages the creation and deletion of the shard and provides reads and writes for the clients, while one or more borrower nodes called “logical-standby” provide only reads after synchronization with the “logical-active” node. In this way, the “logical-active” model achieves low latency close to the “active-standby” model and high throughput close to the “active-active” model.