arXiv subjects

Shengkai Lin

Publications and source records attributed to Shengkai Lin.

7 recordsLinked to original sources

arXiv2026

LiveR: Fine-Grained Elasticity via Live Reconfiguration for Model Training

To reduce user costs and maximize cluster utilization, large model training increasingly leverages volatile but inexpensive GPU capacity, such as spot instances and reclaimable resources in shared clusters. Yet, capitalizing on these economic benefits requires jobs to adapt within the short warning windows that many such environments provide. Existing elastic training systems still treat reconfiguration as stop-and-restart: they externalize distributed state through checkpoints, rebuild the distributed runtime on a new topology, and restart training, turning each resize event into a storage-heavy recovery procedure that incurs substantial downtime from checkpoint I/O, process restart, CUDA initialization, and communicator setup. We present LiveR, a live reconfiguration runtime for elastic LLM training that replaces storage-backed restart with a live, bounded-memory handoff between mixed-parallel training worlds. While the current world continues training, LiveR asynchronously prepares the target world, bootstraps newly added workers in isolation to keep heavyweight initialization off the critical path, and streams model state directly over high-bandwidth interconnects while reshaping it online across tensor, pipeline, and data parallel dimensions. Once the target world is ready, LiveR performs a lightweight commit that switches training to the new configuration without stop-and-restart on the live path. We implement LiveR atop Megatron-LM and PyTorch and evaluate it end-to-end on a multi-node GPU cluster. Across diverse reconfiguration scenarios, LiveR reduces downtime from minutes to seconds, accelerates reconfiguration by 14$\times$-23$\times$ over checkpoint/restart baselines, incurs minimal steady-state overhead, and sustains up to 99% training goodput under volatile-resource conditions, making volatile low-cost GPU capacity far more practical for LLM training.

cs.DC↗

arXiv2026

DynaFlow: Transparent and Flexible Intra-Device Parallelism via Programmable Operator Scheduling

Intra-device parallelism addresses resource under-utilization in ML inference and training by overlapping the execution of operators with different resource usage. However, its wide adoption is hindered by a fundamental conflict with the static, sequential programming model of existing frameworks. Integrating these strategies requires invasive, model-specific code overhauls, representing an intractable engineering cost. This is further amplified by the high sensitivity of strategies to execution contexts (e.g., workload, model architecture, hardware), forcing developers to implement and maintain multiple specialized solutions. To address this, we propose DynaFlow, a framework that enables the transparent and flexible integration of intra-device parallelism by decoupling the logical model definition from the physical execution schedule. DynaFlow introduces a flexible frontend with annotations for graph partitioning and a programmable interface for defining custom intra-device parallelism strategies. Its efficient backend manages complex control/data-flow asynchronously, uses custom memory management to eliminate copy overheads, and preserves compatibility with optimizations like CUDA Graphs and TorchInductor. We demonstrate that DynaFlow can integrate representative parallelism strategies into 6 state-of-the-art ML systems with minimal code changes, achieving up to a 1.29x throughput improvement. DynaFlow is publicly available at https://github.com/uw-syfi/DynaFlow.

cs.DC↗

arXiv2026

Libra: Accelerating Socket I/O via Programmable Selective Data Copying

Layer-7 (L7) proxies are critical to modern cloud-native systems, yet their performance is increasingly bottlenecked by copying entire payloads across the kernel-user boundary. Existing approaches reduce this overhead but typically sacrifice compatibility with unmodified POSIX applications, introduce new APIs, or require specialized environments. We show that, under conventional OS abstractions, fully eliminating kernel-user copies while preserving standard socket semantics for unmodified proxies is fundamentally impossible. This leads to a practical insight: in common L7 workloads, proxies inspect only small metadata (e.g., HTTP headers) for routing, while forwarding the bulk payload unchanged. Based on this insight, we present Libra, an OS-level selective-copy framework that copies only metadata to the user space and retains the bulk payload in the kernel for forwarding, reducing data movement without breaking compatibility. Libra uses eBPF to identify protocol-specific metadata boundaries and coordinate selective copy and payload reuse across receive and transmit paths, all without modifying the socket API. Implemented in Linux and evaluated with unmodified Nginx and HAProxy, Libra improves plaintext throughput by up to 4.2x and reduces P99 tail latency by over 90%. With hardware-offloaded kTLS, it boosts encrypted throughput by 2.0x and cuts tail latency by 65%.

cs.NI↗

arXiv2026

SAIR: Cost-Efficient Multi-Stage ML Pipeline Autoscaling via In-Context Reinforcement Learning

Multi-stage ML inference pipelines are difficult to autoscale due to heterogeneous resources, cross-stage coupling, and dynamic bottleneck migration. We present SAIR, an autoscaling framework that uses an LLM as an in-context reinforcement learning controller, improving its policy online from reward-labeled interaction histories without gradient updates. SAIR combines Pareto-dominance reward shaping with a provable separation margin, surprisal-guided experience retrieval for context efficiency, and fine-grained GPU rate control via user-space CUDA interception. We provide regret analysis decomposing error into retrieval coverage and LLM selection components. On four ML serving pipelines under three workload patterns, SAIR achieves the best or tied-best P99 latency and effective resource cost among deployed baselines, improving P99 by up to 50% and reducing effective cost by up to 97% (under GPU rate-control assumptions), with 86% bottleneck detection accuracy and no offline training.

cs.LG↗

arXiv2025

SHIFT: Exploring the Boundary of RDMA Network Fault Tolerance

Under gang scheduling for large-scale distributed large language model (LLM) training, a single network anomaly can stall or abort an entire job. Current network fault tolerance mechanisms typically adopt a ``fallback and bypass'' approach within the switching fabric and at the access layer, tolerating in-network and access-layer failures. We explore whether RDMA fault tolerance can be extended to the cross-NIC level by failing over traffic to intra-host backup NICs. For the first time, we prove a fundamental Trilemma: it is impossible to have Cross-NIC RDMA failover that simultaneously preserves Exactly-Once Execution, Receiver-NIC Opacity, and a Zero-Copy datapath. Fortunately, we observe that dominant training frameworks (e.g., NCCL) rely on idempotent bulk transfers that tolerate relaxed memory ordering, as long as notification ordering is preserved. Leveraging this insight, we present SHIFT, a user-space RDMA layer that provides cross-NIC fault tolerance while preserving correct memory semantics. We implement SHIFT in \texttt{rdma-core} and evaluate it with PyTorch distributed training. Results show that SHIFT incurs negligible overhead during normal operation and successfully masks fatal NIC failures and link anomalies, allowing training to continue without costly restarts.

cs.NI↗

arXiv2023

Isolated Scheduling for Distributed Training Tasks in GPU Clusters

Distributed machine learning (DML) technology makes it possible to train large neural networks in a reasonable amount of time. Meanwhile, as the computing power grows much faster than network capacity, network communication has gradually become the bottleneck of DML. Current multi-tenant GPU clusters face network contention caused by hash-collision problem which not only further increases the overhead of communication, but also creates unfairness and affects the user experience. In this paper, we firstly analyse how network contention affects the training time in a cluster with 32 NVIDIA V100 GPUs. Then we propose vClos to eliminate network contention by jointly optimizing network topology and communication pattern in distributed training. An OCS-vClos which introduces a layer of optical circuit switches (OCSs) in the leaf-spine network is also proposed to reduce potential network resource fragmentation caused by resource allocation strategy in vClos. Testbed experiments and real-trace-based large-scale simulations are conducted to demonstrate the superiority of vClos over existing network resource scheduling strategies.

cs.DC↗

arXiv2023

ONCache: A Cache-Based Low-Overhead Container Overlay Network

Recent years have witnessed a widespread adoption of containers. While containers simplify and accelerate application development, existing container network technologies either incur significant overhead, which hurts performance for distributed applications, or lose flexibility or compatibility, which hinders the widespread deployment in production. We carefully analyze the kernel data path of an overlay network, quantifying the time consumed by each segment of the data path and identifying the \emph{extra overhead} in an overlay network compared to bare metal. We observe that this extra overhead generates repetitive results among packets, which inspires us to introduce caches within an overlay network. We design and implement ONCache (\textbf{O}verlay \textbf{N}etwork \textbf{Cache}), a cache-based container overlay network, to eliminate the extra overhead while maintaining flexibility and compatibility. We implement ONCache using the extended Berkeley Packet Filter (eBPF) with only 524 lines of code, and integrate it as a plugin of Antrea. With ONCache, containers attain networking performance akin to that of bare metal. Compared to the standard overlay networks, ONCache improves throughput and request-response transaction rate by 12\% and 36\% for TCP (20\% and 34\% for UDP), respectively, while significantly reducing per-packet CPU overhead. Popular distributed applications also benefit from ONCache.

cs.NI↗