VoltanaLLM: Energy-Efficient and SLO-Aware Disaggregated LLM Serving via Adaptive Frequency Control and State-Space Routing

Researchers have introduced two new systems, VoltanaLLM and PARS, aimed at improving the efficiency of Large Language Model (LLM) inference. VoltanaLLM reduces energy consumption by up to 36.3% while maintaining strict latency SLOs^[1].

VoltanaLLM is the first system to explicitly target and reduce energy bloat in modern prefill-decode disaggregated LLM serving. It achieves this by combining phase-specific frequency selection with a decode state-space guided router. According to the researchers, LLM inference exhibits a U-shaped energy-frequency curve, creating 'sweet spots' that depend on phase behavior and load^[1]. Meanwhile, PARS, a prompt-aware LLM task scheduler, has been shown to mitigate Head-of-Line (HOL) blocking in LLM inference tasks, achieving up to 15.7x reduction in latency compared to the vLLM default scheduler^[2]. PARS approximates shortest-job-first scheduling through pairwise ranking with a margin ranking loss and effectively predicts response-length-based task ordering directly from prompts. The system integrates seamlessly with vLLM, a state-of-the-art LLM serving system. Extensive experiments demonstrate that PARS generalizes effectively across diverse LLMs without requiring model-specific retraining^[2]. The introduction of these systems is expected to contribute to more sustainable and scalable LLM deployment.