sLLM: Accelerating LLM Inference using Semantic Load Balancing with Shared Memory Data Structures

As Large Language Models (LLMs) are increasingly deployed to support a broad spectrum of applications, enhancing inference efficiency and minimizing costs have become critical areas of focus. To address these challenges, researchers have explored optimizing the Key-Value (KV) cache within LLMs. However, existing approaches have not considered the potential benefits of sharing KV caches across multiple requests in a cluster environment. Addressing this gap, we introduce sLLM, a novel system that integrates an efficient shared-memory-based Semantic Load Balancer with a KV cache sharing mechanism. This design significantly reduces the need for recomputation during LLM inference, which enhances inference performance. Our evaluation of the sLLM system showcases its effectiveness: the Semantic Load Balancer achieves up to a 7× reduction in latency when dispatching requests, while the system as a whole can decrease the Time-To-First-Token (TTFT) for LLM inferences by 30 - 58%.