arXiv subjects

Zhehu Yuan

Publications and source records attributed to Zhehu Yuan.

4 recordsLinked to original sources

arXiv2024

Heat: Satellite's meat is GPU's poison

In satellite applications, managing thermal conditions is a significant challenge due to the extreme fluctuations in temperature during orbital cycles. One of the solutions is to heat the satellite when it is not exposed to sunlight, which could protect the satellites from extremely low temperatures. However, heat dissipation is necessary for Graphics Processing Units (GPUs) to operate properly and efficiently. In this way, this paper investigates the use of GPU as a means of passive heating in low-earth orbit (LEO) satellites. Our approach uses GPUs to generate heat during the eclipse phase of satellite orbits, substituting traditional heating systems, while the GPUs are also cooled down during this process. The results highlight the potential advantages and limitations of this method, including the cost implications, operational restrictions, and the technical complexity involved. Also, this paper explores the thermal behavior of GPUs under different computational loads, specifically focusing on execution-dominated and FLOP-dominated workloads. Moreover, this paper discusses future directions for improving GPU-based heating solutions, including further cost analysis, system optimization, and practical testing in real satellite missions.

cs.DC↗

arXiv2022

AlphaMLDigger: A Novel Machine Learning Solution to Explore Excess Return on Investment

How to quickly and automatically mine effective information and serve investment decisions has attracted more and more attention from academia and industry. And new challenges have arisen with the global pandemic. This paper proposes a two-phase AlphaMLDigger that effectively finds excessive returns in a highly fluctuated market. In phase 1, a deep sequential natural language processing (NLP) model is proposed to transfer Sina Microblog blogs to market sentiment. In phase 2, the predicted market sentiment is combined with social network indicator features and stock market history features to predict the stock movements with different Machine Learning models and optimizers. The results show that the ensemble models achieve an accuracy of 0.984 and significantly outperform the baseline model. In addition, we find that COVID-19 brings data shift to China's stock market.

q-fin.CP↗

arXiv2022

Forgetful Forests: high performance learning data structures for streaming data under concept drift

Database research can help machine learning performance in many ways. One way is to design better data structures. This paper combines the use of incremental computation and sequential and probabilistic filtering to enable "forgetful" tree-based learning algorithms to cope with concept drift data (i.e., data whose function from input to classification changes over time). The forgetful algorithms described in this paper achieve high time performance while maintaining high quality predictions on streaming data. Specifically, the algorithms are up to 24 times faster than state-of-the-art incremental algorithms with at most a 2% loss of accuracy, or at least twice faster without any loss of accuracy. This makes such structures suitable for high volume streaming applications.

cs.LG↗

arXiv2020

Cost-Effective Data Feeds to Blockchains via Workload-Adaptive Data Replication

Feeding external data to a blockchain, a.k.a. data feed, is an essential task to enable blockchain interoperability and support emerging cross-domain applications, notably stablecoins. Given the data-intensive feeds in real life (e.g., high-frequency price updates) and the high cost in using blockchain, namely Gas, it is imperative to reduce the Gas cost of data feeds. Motivated by the constant-changing workloads in finance and other applications, this work focuses on designing a dynamic, workload-aware approach for cost effectiveness in Gas. This design space is understudied in the existing blockchain research which has so far focused on static data placement. This work presents GRuB, a cost-effective data feed that dynamically replicates data between the blockchain and an off-chain cloud storage. GRuB's data replication is workload-adaptive by monitoring the current workload and making online decisions w.r.t. data replication. A series of online algorithms are proposed that achieve the bounded worst-case cost in blockchain's Gas. GRuB runs the decision-making components on the untrusted cloud off-chain for lower Gas costs, and employs a security protocol to authenticate the data transferred between the blockchain and cloud. The overall GRuB system can autonomously achieve low Gas costs with changing workloads. We built a GRuB prototype functional with Ethereum and Google LevelDB, and supported real applications in stablecoins. Under real workloads collected from the Ethereum contract-call history and mixed workloads of YCSB, we systematically evaluate GRuB's cost which shows a saving of Gas by 10% ~ 74%, with comparison to the baselines of static data-placement.

cs.CR↗