A Controlled Experimental Evaluation of Configuration and Cache Architecture in LLM Serving Engines

Muhammad Faizan Anwar; Sina Keshvadi

Home/Archive/arXiv

arXiv · 2611.05991

A Controlled Experimental Evaluation of Configuration and Cache Architecture in LLM Serving Engines

Muhammad Faizan Anwar·Sina Keshvadi

View original source ↗Read source document ↗Export metadata ↓

Abstract

Multi-agent LLM applications repeat most of their prompt tokens across requests, and serving engines reuse the cached state of shared prefixes. vLLM and SGLang reuse it at different granularities, and benchmarks of their default deployments report large performance gaps, usually credited to the cache. Such comparisons rarely control for other configuration differences. One of them, whether a long prompt is prefilled in chunks, is set by each engine's defaults from the hardware and the release, and it confounds the comparison. We built a reproducible testbed on Kubernetes and a seeded multi-agent workload generator, and compared the two engines under vendor defaults and with chunked prefill matched. SGLang's configuration is the same in both arms, which makes it an internal control across independently provisioned deployments. Under defaults, vLLM's median time to first token is up to 64x SGLang's. Matching one setting removes 78% of the capacity gap where that gap is largest and 30 to 67% of the latency gap. Near saturation the remaining latency gap reflects queue growth. A fluid queue model driven only by a 4.4% throughput shortfall reproduces it, and shows that such a gap scales with the length of the benchmark run. What survives matching is a tail-latency advantage of 2.4 to 6.4x at equal throughput, consistent with cache-aware ordering of waiting requests. We derive six recommendations for comparing serving engines.

Explore related subjects

cs.DC cs.LG cs.PF

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Muhammad Faizan Anwar, Sina Keshvadi. 2026-11-04. A Controlled Experimental Evaluation of Configuration and Cache Architecture in LLM Serving Engines. https://arxiv.org/abs/2611.05991

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

FedDyMem: Efficient Federated Learning with Dynamic Memory and Memory-Reduce for Unsupervised Image Anomaly Detection

Unsupervised image anomaly detection (UAD) has become a critical process in industrial and medical applications, where models are trained on normal samples without requiring labeled anomalies. However, even normal images may contain sensitive product designs, manufacturing information, or institution-specific population characteristics, making centralized data collection difficult in practical deployments. Federated UAD provides a promising decentralized solution, but it remains challenging because clients use only normal training samples, their feature distributions remain highly heterogeneous within the same task, and communication should be efficient while limiting the exposure of sensitive client information. To address these challenges, we propose FedDyMem, an efficient federated learning method with dynamic memory and memory-reduce for unsupervised image anomaly detection. Instead of exchanging full model parameters, FedDyMem facilitates knowledge sharing between the clients and server through reduced dynamic memory banks. In the local clients, a memory generator and a metric loss are employed to improve the consistency of the feature distribution for normal samples, leveraging the local model to update the memory bank dynamically. For efficient communication and data privacy, a memory-reduce method based on weighted averages is proposed to significantly decrease the scale of memory banks. This representation reduces communication overhead and empirically lowers reconstruction fidelity under inversion attacks. On the server, global memory is constructed and distributed to individual clients through k-means aggregation. Experiments on six federated UAD tasks constructed from eleven public industrial and medical datasets show that FedDyMem consistently improves anomaly detection and localization while reducing communication cost.

cs.DC↗

arXiv2026

Twig: Parallel and Memory-Efficient Tree Speculative Decoding for Hybrid-Attention LLMs

Emerging LLMs increasingly interleave full attention with recurrent linear attention to reduce long-context serving costs. Tree speculative decoding could further accelerate these hybrid models, but recurrent linear-attention layers carry parent-dependent states. Existing engines therefore verify proposal trees sequentially and snapshot the full state per node, causing latency and memory to grow with tree and batch sizes. We present Twig, a kernel--runtime co-design for tree speculative decoding of hybrid LLMs. Twig derives a tree-structured factorization of the linear-attention recurrence and realizes it with a resource-efficient GPU kernel that verifies all nodes in parallel, accelerating linear-attention tree verification by 3.4--7.7$\times$. It losslessly represents speculative state updates as token-level factors and reconstructs only the accepted state, reducing transient state memory by 82--99$\times$ and freeing GPU memory for KV caches. We integrate Twig into SGLang, a widely deployed production LLM serving engine, and further optimize its runtime through batch-wide verification budgeting. Across four models, two GPU platforms, and diverse datasets, Twig achieves up to $4.72\times$ the offline decode throughput of autoregressive decoding and up to $2.03\times$ that of the strongest tree-speculative baseline. Under online agent workloads, Twig reduces TTFT and TPOT by up to $67.6\%$ and $49.9\%$, respectively, over the strongest tree-speculative baseline.

cs.DC↗

arXiv2026

Lantern: Finding Committable Transactions via Back-Propagation on DAGs

Existing concurrency control protocols either introduce nondeterminism, resulting in a serial execution-replay dependency between primary and replica nodes, or rely on impractical prior knowledge of transaction read-write sets. In this paper, we present Lantern, a deterministic concurrency control protocol tailored for high-performance transaction processing systems operating without prior knowledge. The key insight of Lantern is that all zero-out-degree transaction vertices in the local dependency graph can be safely committed in ascending order using an overwrite-permissive strategy. We further introduce a novel Back-Propagation mechanism that iteratively propagates dependency states from sink to source vertices to identify additional committable transactions. We also propose Conflict-Free Batch Selection (CFBS) for read-modify-write intensive scenarios. We integrate Lantern into the open-source blockchain platform ChainMaker. Extensive evaluations on YCSB and SmallBank benchmarks demonstrate that Lantern achieves up to a 4.2x throughput speedup over Aria and improves the throughput of ChainMaker's execution layer by at least 2.2x.

cs.DC↗