Aug 2026· Proceedings of the ACM SIGCOMM 2026 Conference· 0 citations· 87 references
TL;DR
This work proposes Turbo, a first-of-its-kind in-network aggregation system that accelerates long-context inference by offloading query broadcast and attention aggregation to switches and introduces a rolling forward scheme that propagates states to enable cross-stage updates.
Abstract
LLM supporting long contexts faces a critical memory bottleneck due to the linear growth of KV cache. Distributing the storage across multiple GPUs alleviates this burden but introduces significant communication overhead or traffic incast, especially during the decoding phase. We propose Turbo, a first-of-its-kind in-network aggregation system that accelerates long-context inference by offloading query broadcast and attention aggregation to switches. We address three key challenges to map complex attention mechanisms onto restricted switch hardware: (i) To bypass the switch's inability to buffer global states or perform complex operations, we devise online table-based aggregation, which decomposes global reduction into pairwise operations and approximates nonlinear functions via lookup tables. (ii) To circumvent the restriction on retroactive state access in RMT pipelines, we introduce a rolling forward scheme that propagates states to enable cross-stage updates. (iii) To mitigate aggregation stragglers caused by topology-induced load imbalance, we construct a load-aware aggregation tree that optimizes workload distribution. Evaluations on a Tofino2-based testbed show that Turbo reduces end-to-end inference latency by up to 37%. Large-scale simulations on NS-3 demonstrate that Turbo significantly outperforms state-of-the-art baselines in both inference latency and network traffic reduction with negligible accuracy loss.
Modern large language model (LLM) inference suffers from severe Time-To-First-Token (TTFT) bottlenecks. Existing prefix KV caching mechanisms are inherently stateless, forcing a trade-off between cross-chunk attention accuracy and online recomputation overhead. To address this issue, we propose Pegasus, a novel stateful prefix KV caching system that aims to achieve full-context attention accuracy while avoiding costly recomputation. To handle the exponential growth of context states under limited memory capacity, Pegasus employs a Recursive Path-Pruning Caching (RPPC) algorithm to selectively cache high-value states based on access frequency, memory footprint, and asymmetric latency benefit. In addition, Pegasus introduces a transition-based KV management mechanism to mitigate cache-miss overhead. By exploiting the sparsity of state-dependent KV variations, it replaces expensive attention recomputation and I/O-intensive tensor reloading with lightweight sparse state transitions. Extensive experiments show that Pegasus improves end-to-end serving throughput by 45.9% on average, reduces TTFT by up to 78.5%, and lowers cache-miss recovery overhead by more than 72%.
Fahao Chen, Peng Li, Dongxiao Yu et al.· Fall Joint Computer Conferen...· 0 citations
With the rapid advancement of deep learning technology, the parameter scale of large language models has grown exponentially, expanding from hundreds of millions in the early stages to hundreds of billions or even trillions today. Pipeline inference is a crucial approach enabling efficient inference in large language models. However, existing pipeline inference relies on static layer allocation, ignoring the intrinsic variance in layer-wise computation and memory footprints, as well as runtime fluctuations in request rates and sequence lengths. Consequently, under dynamic workloads, compute-dense stages rapidly bottleneck the pipeline and induce severe queue blocking while leaving other devices idle, ultimately degrading end-to-end latency and severe GPU underutilization. To address these challenges, this paper proposes a dynamic pipeline parallel inference algorithm. Centering on the three phases of LLM pipeline inference—partitioning, updating, and migration—the algorithm introduces: (1) A proactive update trigger mechanism driven by multidimensional load forecasting. Rather than relying on reactive bottleneck indicators, it translates projected request rates and token lengths into stage-level VRAM demands, preemptively initiating reconfiguration only when impending hardware capacity violations are detected; (2) A joint partition-migration optimization strategy utilizing a two-stage biased random key genetic algorithm. By embedding a maximum-weight bipartite matching formulation into the evolutionary fitness evaluation, this strategy mathematically couples pipeline boundary search with physical state mapping, maximizing resident parameter reuse to guarantee minimal-overhead model migration; (3) Distributed cluster experiments conducted using public datasets and the Ray framework demonstrate that the proposed method outperforms existing state-of-the-art pipeline inference solutions in metrics including response latency and resource overhead, specifically improving throughput by 2.3% compared to the SOTA framework.
To accelerate large language model (LLM) inference, pipeline parallelism partitions model layers into sequential stages, each assigned to a different device for concurrent execution. However, this method often suffers from pipeline bubbles caused by imbalanced computation in the tail stage. While upstream stages focus solely on layer-forward operations, the final stage must also handle additional post-processing tasks like sampling, which introduces significant latency. This discrepancy in workload leads to pipeline misalignment, forcing upstream stages to idle and degrading overall performance. Existing frameworks typically distribute layers evenly across stages without accounting for computational load differences. To address this, we propose DynaPipe , a dynamic layer redistribution scheme that adaptively balances computation by predicting execution latency in real time. Moreover, we introduce an asynchronous key-value (KV) cache migration coordinator to enable non-blocking layer redistribution during inference. Experiments on representative LLMs demonstrate that DynaPipe reduces average end-to-end request latency by 8% to 41% across diverse workloads, outperforming state-of-the-art pipeline parallelism systems. Our implementation is publicly available at https://github.com/xhx1022/DynaPipe .
Or OrionInfer, an adaptive LLM serving system that aligns inference strategies with real-time demand and introduces three key techniques: runtime switching between data parallelism and tensor parallelism with negligible overhead, an efficient inference pipeline that preserves batching efficiency during parallelism transitions, and live-migration-based load balancing to alleviate memory pressure and improve resource utilization.
Jingqi Feng, Guang Yang, Yukai Huang et al.· Proceedings of the 32nd ACM...· 0 citations
Large language model (LLM) agents repeatedly process long, privacy-sensitive contexts, while cloud-only deployment exposes user data beyond the trusted endpoint and fully local deployment often requires costly hardware. Split inference offers a middle ground by executing the model head, tail, and tools locally and the middle layers in the cloud, but its local-cloud-local path transfers boundary hidden states at every invocation and creates a critical communication bottleneck. We present \system, a reference-aware codec that retrieves exact-token historical spans for prefill uplinks, reuses the reconstructed uplink state for same-round prefill downlinks, and generates boundary-specific decode references with lightweight causal predictors. RAC applies grouped affine alignment and calibrated residual quantization with optional prefill outliers, while sender-side wire-format reconstruction synchronizes subsequent references and offline calibration accounts for quality and packed representation costs. Across three models and nine evaluated model-link pairs, Raw-to-RAC mean time to first token (TTFT) and time per output token (TPOT) ratios are 1.24-2.72$\times$ and 1.01-2.79$\times$, while the 12 non-perplexity task-score changes range from $-0.40$ to $+2.50$ points.
Guotao Yang, M. Zhao, Haopeng Li et al.· 0 citations
DualPath is an inference system that breaks this bottleneck by introducing dual-path KV-Cache loading and enables a novel storage-to-decode path, in which the KV-Cache is loaded into decoding engines and then efficiently transferred to prefill engines via RDMA over the compute network.
Yongtong Wu, Shaoyuan Chen, Rilin Huang et al.· Proceedings of the ACM SIGCO...· 0 citations
