arXiv subjects

Wenquan Yang

Publications and source records attributed to Wenquan Yang.

2 recordsLinked to original sources

arXiv2026

Hotspot-Aware Scheduling of Virtual Machines with Overcommitment for Ultimate Utilization in Cloud Datacenters

We address the problem of under-utilization of resources in datacenters during cloud operations, specifically focusing on the challenge of online virtual machine (VM) scheduling. Rather than following the traditional approach of scheduling VMs based solely on their static flavors, we take into account their dynamic CPU utilization. We employ $Γ$-robustness theory to manage the dynamic nature and introduce a novel variant of bin packing - Probabilistic k-Bins Packing (PkBP), which theoretically protects the Physical Machines (PMs) from hotspots formation within a specified probability $α$. We develop a scheduling algroithm named CloseRadiusFit and cold-start AI based prediction algorithms for the online version of PkBP. To verify the quality of our approach towards the optimal solutions, we solve the Offline PkBP problem by designing a novel Mixed Integer Linear Programming (MILP) model and a combination of numerical upper and lower bounds. Our experimental results demonstrate that CloseRadiusFit achieves narrow gaps of 1.6% and 3.1% when compared to the lower and upper bounds, respectively.

cs.DC↗

arXiv2026

Huawei Cloud Model-as-a-Service on the CloudMatrix384 SuperPod

Scaled-out MoE LLMs and scaled-up SuperPods create new systems challenges for production Model-as-a-Service (MaaS), requiring disaggregation, low-latency communication, and decentralized serving. This report presents xDeepServe, the production serving system behind Huawei Cloud's MaaS offering on CloudMatrix384, a 48-server SuperPod with 384 Ascend 910C chips connected by a high-bandwidth UB fabric and global shared memory. It serves models including DeepSeek, Kimi, GLM, Qwen, and MiniMax, among others. xDeepServe is built around Transformerless, a disaggregated execution architecture that decomposes transformer inference into modular units -- attention, feedforward, and MoE -- and supports disaggregated Prefill-Decode and MoE-Attention deployments. To enable disaggregation, we develop XCCL, a memory-semantic communication layer providing microsecond-level point-to-point and scalable all-to-all primitives, and we extend FlowServe with decentralized DP groups and techniques to mitigate stragglers and synchronization variance. In a peak decoding configuration, xDeepServe reaches 2400 tokens/s per Ascend 910C chip at ~50ms time-per-output-token (TPOT).

cs.DC↗