arXiv subjects

Tianxing Sun

Publications and source records attributed to Tianxing Sun.

4 recordsLinked to original sources

arXiv2026

NestPipe: Large-Scale Recommendation Training on 1,500+ Accelerators via Nested Pipelining

Modern recommendation models have increased to trillions of parameters. As cluster scales expand to O(1k), distributed training bottlenecks shift from computation and memory to data movement, especially lookup and communication latency associated with embeddings. Existing solutions either optimize only one bottleneck or improve throughput by sacrificing training consistency. This paper presents NestPipe, a large-scale decentralized embedding training framework that tackles both bottlenecks while preserving synchronous training semantics. NestPipe exploits two hierarchical sparse parallelism opportunities through nested pipelining. At the inter-batch level, Dual-Buffer Pipelining (DBP) constructs a staleness-free five-stage pipeline through dual-buffer synchronization, mitigating lookup bottlenecks without embedding staleness. At the intra-batch level, we identify the embedding freezing phenomenon, which inspires Frozen-Window Pipelining (FWP) to overlap All2All communication with dense computation via coordinated stream scheduling and key-centric sample clustering. Experiments on production GPU and NPU clusters with 1,536 workers demonstrate that NestPipe achieves up to 3.06x speedup and 94.07% scaling efficiency.

cs.DC↗

arXiv2025

Mechanical properties of B12-based orthorhombic metal carboborides. A first principle study

Structural and mechanical properties of B12-based orthorhombic metal carboborides are studied on the basis of first principle DFT approach. The simulations predict the existence of a new family of phases of the composition MeC2B12 (Me=Mg, Ca, Sr, Sc, Y) with similar structure and space symmetry Imma. It has been found that the predicted phases are thermally (dynamically) stable and have considerably better mechanical properties as compared to the reference compound AlMgB14. The respective calculated isotropic elastic moduli and Vickers hardness are significantly higher (G~ 230-250, E~530-550, and Hv~35-55 GPa). These conclusions were confirmed by direct calculations of shear strength for the above phases, which demonstrated the increase of 30-50% in different directions. The enhanced mechanical characteristics of the MgC2B12 -based phases make them promising for creating novel superhard materials

cond-mat.mtrl-sci↗

arXiv2024

Orthorhombic metal carbide-borides MeC$_2$B$_{12}$ (Me=Mg, Ca, Sr) from first principles: structure, stability and mechanical properties

First principle DFT simulations are employed to study structural and mechanical properties of orthorhombic B12-based metal carbide-borides. The simulations predict the existence of Ca- and Sr- based phases with the structure similar to that of experimentally observed earlier compound MeC$_2$B$_{12}$. Dynamical stability of both phases is demonstrated, and the phase MeC$_2$B$_{12}$ is found to be thermodynamically stable. According to simulations, Ca- and Sr- based phases have significantly enhanced mechanical characteristics, which suggest their potential application as superhard materials. Calculated shear and Young moduli of these phases are nearly 250 and 540 GPa, respectively, and estimated Vickers hardness is 45-55 GPa.

cond-mat.mtrl-sci↗

arXiv2021

Energy Management Strategy for Unmanned Tracked Vehicles Based on Local Speed Planning

The hybrid electric system has good potential for unmanned tracked vehicles due to its excellent power and economy. Due to unmanned tracked vehicles have no traditional driving devices, and the driving cycle is uncertain, it brings new challenges to conventional energy management strategies. This paper proposes a novel energy management strategy for unmanned tracked vehicles based on local speed planning. The contributions are threefold. Firstly, a local speed planning algorithm is adopted for the input of driving cycle prediction to avoid the dependence of traditional vehicles on driver's operation. Secondly, a prediction model based on Convolutional Neural Networks and Long Short-Term Memory (CNN-LSTM) is proposed, which is used to process both the planned and the historical velocity series to improve the prediction accuracy. Finally, based on the prediction results, the model predictive control algorithm is used to realize the real-time optimization of energy management. The validity of the method is verified by simulation using collected data from actual field experiments of our unmanned tracked vehicle. Compared with multi-step neural networks, the prediction model based on CNN-LSTM improves the prediction accuracy by 20%. Compared with the traditional regular energy management strategy, the energy management strategy based on model predictive control reduces fuel consumption by 7%.

eess.SY↗