Tag
A survey paper examining forty years of just-in-time compilation research from 1960 to 2000, covering the motivation, design constraints, and a classification scheme for JIT systems.
WAKMan is a workload-aware KV cache manager for LLM servers that automatically selects its eviction policy from live traffic, improving hit rate over vLLM's LRU by 5 percentage points on average across mixed single-turn, multi-turn, and agentic request workloads.
TAO 提出一个工具感知的联合调度框架，用于基于 LLM 的智能体服务，通过动态规划在容量约束下选择和放置会话，相比最优基线最高降低 26.8% 延迟、36.3% TTFT 并提升 29.5% 解码吞吐。
OpenBSD 8.0, the 61st OpenBSD release, was announced on October 20, 2026 with platform improvements for amd64, arm64, riscv64 and sparc64, including larger kernel virtual address space, hibernate support, RISC-V vector/Advanced Interrupt Architecture support, and a switch to the LLVM toolchain on sparc64.
The paper presents Apex, a CPU-GPU collaborative MoE inference system that adaptively assigns GPU-missed experts to either CPU or GPU based on token load, improving prefill speed by up to 51.9% and decode throughput by up to 39.5% over the strongest baseline.
该论文提出ORL（On-policy Residual Learning）框架，通过熵收缩目标与异步残差捕获等系统级协同设计，使扩散语言模型（DLM）在相近质量下将吞吐量提升至约6.9×，在单请求解码速度上超越自回归模型的推测解码方案。
该论文研究在图数据库中直接进行图神经网络的训练与推理，展示使用Cypher执行采样、特征检索和完整前向传播的可行性，优化后训练速度提升约2.5倍，推理速度比冷加载图到内存快10倍以上。
ByteDance and Fudan University researchers present ByteExpress, a software communication layer over RDMA for large-scale parameter-server inference and training, reporting up to 74% lower P99 push latency and significant CPU savings versus TCP in production deployments.
The research paper "B-Trees Are Back: Engineering Fast and Pageable Node Layouts" explores how to design faster, pageable B-tree node layouts to improve database index performance on modern hardware.
MemMux is a local runtime that treats resource governance for fleets of parallel coding agents as runtime verification, providing per-agent memory attribution, complete reclamation, escaped-process visibility, and bounded footprint. Benchmarks show it keeps a 7.5 GiB budget under zero swap and detects all injected escaped processes, with an open-sourced engine and benchmark.
TomasuLLM 是一个将乱序推测执行思想引入 LLM Agent 的运行时，让工具调用在隔离的写时复制沙箱中乱序执行并按序提交验证，从而在 SWE-bench Verified、Terminal-Bench 2.0 和 SWE-Marathon 上分别提升 1.31x、1.35x 和 1.27x 的性能，且在 4,010 条提交验证记录中零误报。
SlideDP presents an asynchronous data-parallel runtime for host-resident layer streaming that fine-tunes LLMs beyond GPU memory across multiple GPUs, achieving 1.46–2.64x throughput over SlideFormer, MegaTrain, and ZeRO-Offload and even exceeding FSDP2 peak throughput while fine-tuning Qwen2.5-72B on four RTX 4090s.
The MLSys 2027 call for papers is announced, with Song Han serving as PC Co-Chair, inviting research and industrial track submissions on machine learning and systems.
The article questions whether the proliferation of AI agents is leading to increased complexity in management and maintenance, potentially outweighing their benefits.
Andrej Karpathy's Stanford lecture offers a concise explanation of AI engineering, focusing on building systems around models with context, memory, tools, and feedback loops.
InfraBench is a benchmark suite for evaluating AI agents on realistic infrastructure management tasks across the full stack, lifecycle, and risk levels. Experiments show that even the strongest agent configurations achieve only 40-88% effective scores and struggle with consistency and safety.
Discussion on Hacker News about whether computer systems are prepared for the first negative leap second, with comments debating the risks and alternatives like leap minutes.
A cheatsheet covering memory segmentation concepts, likely useful for students and developers.
A paper titled 'The Hitchhiker's Guide to Agentic AI: From Foundations to Systems' provides comprehensive textbook-like resources on the basics of LLMs and agentic AI.
A tweet sharing Brendan Gregg's curated list of recommended books covering systems performance, language performance, database performance, and other computer science domains.