Esha Choukse

Pratyush Patel, Esha Choukse, Chaojie Zhang et al.

Recent innovation in large language models (LLMs), and their myriad use-cases have rapidly driven up the compute capacity demand for datacenter GPUs. Several cloud providers and other enterprises have made substantial plans of growth in their datacenters to support these new workloads. One of the key bottleneck resources in datacenters is power, and given the increasing model sizes of LLMs, they are becoming increasingly power intensive. In this paper, we show that there is a significant opportunity to oversubscribe power in LLM clusters. Power oversubscription improves the power efficiency of these datacenters, allowing more deployable servers per datacenter, and reduces the deployment time, since building new datacenters is slow. We extensively characterize the power consumption patterns of a variety of LLMs and their configurations. We identify the differences between the inference and training power consumption patterns. Based on our analysis of these LLMs, we claim that the average and peak power utilization in LLM clusters for inference should not be very high. Our deductions align with the data from production LLM clusters, revealing that inference workloads offer substantial headroom for power oversubscription. However, the stringent set of telemetry and controls that GPUs offer in a virtualized environment, makes it challenging to have a reliable and robust power oversubscription mechanism. We propose POLCA, our framework for power oversubscription that is robust, reliable, and readily deployable for GPU clusters. Using open-source models to replicate the power patterns observed in production, we simulate POLCA and demonstrate that we can deploy 30% more servers in the same GPU cluster for inference, with minimal performance loss

Amey Agrawal, Haoran Qiu, Junda Chen et al. · gatech

Deploying million-token Large Language Models (LLMs) is challenging because production workloads are highly heterogeneous, mixing short queries and long documents. This heterogeneity, combined with the quadratic complexity of attention, creates severe convoy effects where long-running requests stall short, interactive ones, degrading system responsiveness. We present Medha, a serving system that eliminates these convoys by introducing fine-grained, preemptive scheduling to LLM inference. Medha makes preemption practical with a co-designed set of mechanisms -- including Adaptive Chunking and Stream Pipeline Parallel that overcome the perceived inefficiencies and scaling challenges of chunking. Additionally, we present a new parallelism strategy KV-Cache Parallelism to reduce the decode latency and afford interactivity despite very long context. These mechanisms are orchestrated by a Length-Aware Relative Slack (LARS) scheduler, a deadline and heterogeneity-aware scheduling policy that prevents both the convoy effect and the starvation that plagues simpler policies. Under a heterogeneous workload, Medha improves throughput by 5.7x while reducing median and 99th percentile latency by 30x and 174x, respectively, compared to state-of-the-art non-preemptive systems.

Jovan Stojkovic, Chaojie Zhang, Íñigo Goiri et al.

The rapid evolution and widespread adoption of generative large language models (LLMs) have made them a pivotal workload in various applications. Today, LLM inference clusters receive a large number of queries with strict Service Level Objectives (SLOs). To achieve the desired performance, these models execute on power-hungry GPUs causing the inference clusters to consume large amount of energy and, consequently, result in excessive carbon emissions. Fortunately, we find that there is a great opportunity to exploit the heterogeneity in inference compute properties and fluctuations in inference workloads, to significantly improve energy-efficiency. However, such a diverse and dynamic environment creates a large search-space where different system configurations (e.g., number of instances, model parallelism, and GPU frequency) translate into different energy-performance trade-offs. To address these challenges, we propose DynamoLLM, the first energy-management framework for LLM inference environments. DynamoLLM automatically and dynamically reconfigures the inference cluster to optimize for energy and cost of LLM serving under the service's performance SLOs. We show that at a service-level, DynamoLLM conserves 53% energy and 38% operational carbon emissions, and reduces 61% cost to the customer, while meeting the latency SLOs.

Theo Gregersen, Pratyush Patel, Esha Choukse

GPUs are known to be power-hungry, and due to the boom in artificial intelligence, they are currently the major contributors to the high power demands of upcoming datacenters. Most GPU usage in these popular workloads consist of large general matrix-matrix multiplications (GEMMs), which have therefore been optimized to achieve high utilization of hardware resources. In this work, we show that modifying the input data to GEMMs, while maintaining the matrix shapes and sizes can notably change the power consumption of these kernels. We experiment with four kinds of input variations: value distribution, bit similarity, placement, and sparsity, across different data types. Our findings indicate that these variations can change the GPU power usage during GEMM by almost 40%. We hypothesize that input-dependent power usage variations occur due to changes in the number of bit flips in the GPUs. We propose leveraging this property through compiler and scheduler optimizations to manage power and reduce energy consumption.

Melissa Z. Pan, Negar Arabzadeh, Mathew Jacob et al.

Modern retrieval agents expose many configuration choices -- LLM, retriever, number of documents, number of hops, and synthesis strategy -- each shaping both answer quality and serving cost. Today, these pipelines are typically hand-tuned once per workload, leaving substantial per-query optimization untapped. We formulate the problem: given a natural-language query and either an accuracy or a budget target, select from a predefined pipeline catalog the configuration that minimizes cost or maximizes accuracy at inference time. We propose **BRANE**, which uses an LLM to convert each query into workload-specific characteristics, then trains a lightweight per-configuration predictor that estimates whether the pipeline will answer the query correctly. At inference time, **BRANE** selects the configuration that maximizes predicted correctness penalized by cost, exposing a tunable cost-quality tradeoff without retraining. Across MuSiQue, BrowseComp-Plus, and FinanceBench, **BRANE** consistently pushes the cost-quality Pareto frontier, matches the best fixed configuration's accuracy at up to 89% lower cost, and outperforms LLM-routing, rule-based, and fine-tuned Qwen3-4B baselines. These results show that per-query configuration of the full retrieval pipeline is a practical alternative to static workload-level tuning.

Haoran Qiu, Gohar Irfan Chaudhry, Chaojie Zhang et al.

Advances in multi-modal generative models are enabling new applications, from storytelling to automated media synthesis. Most current workloads generate simple outputs (e.g., image generation from a prompt) in batch mode, often requiring several seconds even for basic results. Serving real-time multi-modal workflows at scale is costly and complex, requiring efficient coordination of diverse models (each with unique resource needs) across language, audio, image, and video, all under strict latency and resource constraints. We tackle these challenges through the lens of real-time podcast video generation, integrating LLMs, text-to-speech, and video-audio generation. To meet tight SLOs, we design an adaptive, modular serving system, StreamWise, that dynamically manages quality (e.g., resolution, sharpness), model/content parallelism, and resource-aware scheduling. We leverage heterogeneous hardware to maximize responsiveness and efficiency. For example, the system can lower video resolution and allocate more resources to early scenes. We quantify the trade-offs between latency, cost, and quality. The cheapest setup generates a 10-minute podcast video on A100 GPUs in 1.4 hours (8.4x slower than the real-time) for less than \$25. StreamWise enables high-quality real-time streaming with a sub-second startup delay under $45.

Yueying Li, Yuanfan Chen, Jiayang Chen et al.

LLM serving exhibits extreme length variability, making size-based scheduling difficult in practice. Recent LLM schedulers approximate SJF/SRPT using predicted decode lengths or ranks and primarily report mean-centric metrics such as TTFT and TBT. We show that these prediction-driven policies can be fragile under distribution shifts, bursty arrivals, and GPU memory pressure, while offering limited control over the tail latency (P90-P99) that dominates user experience, even with perfect decode-length knowledge. We introduce a distribution-aware, prediction-free scheduling framework that replaces explicit length prediction with soft priority boosting driven by lightweight statistical signals. Our design co-optimizes scheduling and cache-aware preemption to account for memory-coupled decode dynamics across workload mixes. Evaluated on production and open-source traces, our method reduces P99 TTLT by up to 35-50% relative to SRPT with perfect length knowledge and reduces TTFT by 34-47% across workloads, including reasoning-heavy and chat-heavy tasks. These results demonstrate a robust alternative for optimizing tail latency in online LLM serving.

Haoran Qiu, Anish Biswas, Zihan Zhao et al.

Large multimodal models (LMMs) demonstrate impressive capabilities in understanding images, videos, and audio beyond text. However, efficiently serving LMMs in production environments poses significant challenges due to their complex architectures and heterogeneous characteristics across their multi-stage inference pipelines. We present the first comprehensive systems analysis of two prominent LMM architectures, decoder-only and cross-attention, across six representative open-source models, revealing key systems design implications. We also present an in-depth analysis of production LMM inference traces, uncovering unique workload characteristics, including variable, heavy-tailed request distributions and bursty traffic patterns. Based on these insights, we propose ModServe, a modular LMM serving system that decouples stages for independent optimization and adaptive scaling. ModServe dynamically reconfigures stages and handles bursty traffic with modality-aware scheduling and autoscaling to meet tail latency SLOs while minimizing costs. ModServe achieves 3.3-5.5x higher throughput (leading to 25-41.3% cost saving) while meeting SLOs on a 128-GPU cluster with production traces.

Jovan Stojkovic, Esha Choukse, Chaojie Zhang et al.

With the ubiquitous use of modern large language models (LLMs) across industries, the inference serving for these models is ever expanding. Given the high compute and memory requirements of modern LLMs, more and more top-of-the-line GPUs are being deployed to serve these models. Energy availability has come to the forefront as the biggest challenge for data center expansion to serve these models. In this paper, we present the trade-offs brought up by making energy efficiency the primary goal of LLM serving under performance SLOs. We show that depending on the inputs, the model, and the service-level agreements, there are several knobs available to the LLM inference provider to use for being energy efficient. We characterize the impact of these knobs on the latency, throughput, as well as the energy. By exploring these trade-offs, we offer valuable insights into optimizing energy usage without compromising on performance, thereby paving the way for sustainable and cost-effective LLM deployment in data center environments.

Jovan Stojkovic, Chaojie Zhang, Íñigo Goiri et al.

The rising demand for generative large language models (LLMs) poses challenges for thermal and power management in cloud datacenters. Traditional techniques often are inadequate for LLM inference due to the fine-grained, millisecond-scale execution phases, each with distinct performance, thermal, and power profiles. Additionally, LLM inference workloads are sensitive to various configuration parameters (e.g., model parallelism, size, and quantization) that involve trade-offs between performance, temperature, power, and output quality. Moreover, clouds often co-locate SaaS and IaaS workloads, each with different levels of visibility and flexibility. We propose TAPAS, a thermal- and power-aware framework designed for LLM inference clusters in the cloud. TAPAS enhances cooling and power oversubscription capabilities, reducing the total cost of ownership (TCO) while effectively handling emergencies (e.g., cooling and power failures). The system leverages historical temperature and power data, along with the adaptability of SaaS workloads, to: (1) efficiently place new GPU workload VMs within cooling and power constraints, (2) route LLM inference requests across SaaS VMs, and (3) reconfigure SaaS VMs to manage load spikes and emergency situations. Our evaluation on a large GPU cluster demonstrates significant reductions in thermal and power throttling events, boosting system efficiency.

Gohar Irfan Chaudhry, Esha Choukse, Íñigo Goiri et al.

Compound AI Systems, integrating multiple interacting components like models, retrievers, and external tools, have emerged as essential for addressing complex AI tasks. However, current implementations suffer from inefficient resource utilization due to tight coupling between application logic and execution details, a disconnect between orchestration and resource management layers, and the perceived exclusiveness between efficiency and quality. We propose a vision for resource-efficient Compound AI Systems through a declarative workflow programming model and an adaptive runtime system for dynamic scheduling and resource-aware decision-making. Decoupling application logic from low-level details exposes levers for the runtime to flexibly configure the execution environment and resources, without compromising on quality. Enabling collaboration between the workflow orchestration and cluster manager enables higher efficiency through better scheduling and resource management. We are building a prototype system, called Murakkab, to realize this vision. Our preliminary evaluation demonstrates speedups up to $\sim 3.4\times$ in workflow completion times while delivering $\sim 4.5\times$ higher energy efficiency, showing promise in optimizing resources and advancing AI system design.

Esha Choukse, Brijesh Warrier, Scot Heath et al.

Large Artificial Intelligence (AI) training workloads spanning several tens of thousands of GPUs present unique power management challenges. These arise due to the high variability in power consumption during the training. Given the synchronous nature of these jobs, during every iteration there is a computation-heavy phase, where each GPU works on the local data, and a communication-heavy phase where all the GPUs synchronize on the data. Because compute-heavy phases require much more power than communication phases, large power swings occur. The amplitude of these power swings is ever increasing with the increase in the size of training jobs. An even bigger challenge arises from the frequency spectrum of these power swings which, if harmonized with critical frequencies of utilities, can cause physical damage to the power grid infrastructure. Therefore, to continue scaling AI training workloads safely, we need to stabilize the power of such workloads. This paper introduces the challenge with production data and explores innovative solutions across the stack: software, GPU hardware, and datacenter infrastructure. We present the pros and cons of each of these approaches and finally present a multi-pronged approach to solving the challenge. The proposed solutions are rigorously tested using a combination of real hardware and Microsoft's in-house cloud power simulator, providing critical insights into the efficacy of these interventions under real-world conditions.

Colin Hong, Xu Guo, Anand Chaanan Singh et al.

Recently, Test-Time Scaling (TTS) has gained increasing attention for improving LLM reasoning performance at test time without retraining the model. A notable TTS technique is Self-Consistency (SC), which generates multiple reasoning chains in parallel and selects the final answer via majority voting. While effective, the order-of-magnitude computational overhead limits its broad deployment. Prior attempts to accelerate SC mainly rely on model-based confidence scores or heuristics with limited empirical support. For the first time, we theoretically and empirically analyze the inefficiencies of SC and reveal actionable opportunities for improvement. Building on these insights, we propose Slim-SC, a step-wise pruning strategy that identifies and removes redundant chains using inter-chain similarity at the thought level. Experiments on three STEM reasoning datasets and two recent LLM architectures show that Slim-SC reduces inference latency and KVC usage by up to 45% and 26%, respectively, with R1-Distill, while maintaining or improving accuracy, thus offering a simple yet efficient TTS alternative for SC.