~ similar to 2606.00866· 20 results
Chaokun Chang, Yukun Zhou, Kaihua Fu, Dakai An +9 more
This paper introduces AgentSysBench, a benchmark suite and measurement toolkit for agentic applications, and identifies six properties that distinguish agentic workloads from conventional LLM serving.
Shubham Tiwari, Tapan Chugh, Nash Rickert, Simon Peter +2 more
This paper presents a new KVCache management layer called CacheWise that improves performance for coding agent workloads.
PIMID is an execution- and trace-driven full-system simulator for Processing-in-Memory systems, supporting multiple memory technologies, execution models, and placement of processing elements.
This paper argues that idle cores on contemporary multicore processors can return compute capacity and proposes a budget-centric view of service core systems.
This paper identifies suboptimalities in default GPU-first scheduling for AI tool workloads and proposes an agentic scheduler that adaptively assigns tools to GPU or CPU based on runtime factors.
Shuhuai Li, Shengdong Ye, Dongdong Ge, Yinyu Ye +1 more
This paper presents TAO, a tool-aware joint scheduling framework for LLM-based agentic serving that reduces latency, mean time to first token, and increases decode throughput.
Ya Zhang, Tong Lei, Yao Chen, Yonggang Che +3 more
This paper introduces CW-Ghost, a method for estimating cache line fill volume and determining helper-thread prefetching granularity based on cache capacity constraints.
Haozhe Wang, Lina Chen, Fanqi Wan, Qi Han +4 more
This paper presents Metal Sandbox, a system for efficient on-device inference on macOS that combines restricted file access, native command execution, and command-level scheduling.
Zhuoren Ye, Tianyu Wo, Dinghao Xue, Mingming Zhang +3 more
This paper proposes CrossPool, a serving engine for cold Machine Learning Models (LLMs) that separates weights and KV-cache into two GPU memory pools to improve GPU memory utilization and long-context…
This paper proposes StickyInvoc, a task model that reduces overhead in high-throughput workflows by decoupling state creation and destruction for LLM inference tasks.
Combines cross-framework workload characterization with framework-aware resource controller for co-located agents
This paper studies software aging in GPU-based LLM serving systems using a 216-hour campaign.
This paper presents a measurement protocol and benchmarking harness for evaluating the performance of different multi-node strategies for submitting tasks on shared Slurm clusters, focusing on their i…
The paper characterizes 'dead-entry' TLB misses in GPUs, which occur when recently evicted translations are immediately re-walked, and proposes DEPOT, a Bloom filter mechanism that significantly reduc…
Abraham Gonzalez, Raghav Gupta, Akanksha Jain, Hanna Alam +11 more
This paper presents ArchAgent v2, a framework that scales automated microarchitecture search to multi-level data prefetching using a cascaded evolutionary search and hardware-realizability feedback lo…
Leonid Kondrashov, Hongrui Liu, JooYoung Park, Boxi Zhou +13 more
Aries is a full-stack experimentation framework for autonomous agents that separates task semantics from execution configurations, reconstructs agent trajectories, and exposes stateful tool execution.