Researchers have developed JigSLO, a novel system designed to optimize the serving of multiple large language models (LLMs) on shared graphics processing units (GPUs). JigSLO addresses memory bandwidth contention by treating prefill and decode phases of LLMs as separate instances, allowing for complementary co-location. This approach estimates latency based on contended and uncontended periods, enabling the system to maximize throughput while meeting latency targets. In evaluations across chat, code generation, and summarization tasks, JigSLO demonstrated significant improvements, achieving 2.3x better SLO attainment and 1.2x higher throughput compared to existing methods, while using half the GPUs for long-context workloads. AI
IMPACT Optimizes LLM serving efficiency, potentially reducing infrastructure costs and improving response times for multi-LLM deployments.
RANK_REASON The cluster contains an academic paper detailing a new system for LLM serving. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- DagsHub
- Gotit.pub
- graphics processing unit
- Hugging Face
- JigSLO
- Kyungmin Bin
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →