Jayashree Mohan

Amey Agrawal, Anmol Agarwal, Nitin Kedia et al. · gatech

Serving large language models (LLMs) in production can incur substantial costs, which has prompted recent advances in inference system optimizations. Today, these systems are evaluated against conventional latency and throughput metrics (eg. TTFT, TBT, Normalised Latency and TPOT). However, these metrics fail to fully capture the nuances of LLM inference, leading to an incomplete assessment of user-facing performance crucial for real-time applications such as chat and translation. In this paper, we first identify the pitfalls of current performance metrics in evaluating LLM inference systems. We then propose Etalon, a comprehensive performance evaluation framework that includes fluidity-index -- a novel metric designed to reflect the intricacies of the LLM inference process and its impact on real-time user experience. Finally, we evaluate various existing open-source platforms and model-as-a-service offerings using Etalon, discussing their strengths and weaknesses. Etalon is available at https://github.com/project-etalon/etalon.

Amey Agrawal, Ashish Panwar, Jayashree Mohan et al. · gatech

Large Language Model (LLM) inference consists of two distinct phases - prefill phase which processes the input prompt and decode phase which generates output tokens autoregressively. While the prefill phase effectively saturates GPU compute at small batch sizes, the decode phase results in low compute utilization as it generates one token at a time per request. The varying prefill and decode times also lead to imbalance across micro-batches when using pipeline parallelism, resulting in further inefficiency due to bubbles. We present SARATHI to address these challenges. SARATHI employs chunked-prefills, which splits a prefill request into equal sized chunks, and decode-maximal batching, which constructs a batch using a single prefill chunk and populates the remaining slots with decodes. During inference, the prefill chunk saturates GPU compute, while the decode requests 'piggyback' and cost up to an order of magnitude less compared to a decode-only batch. Chunked-prefills allows constructing multiple decode-maximal batches from a single prefill request, maximizing coverage of decodes that can piggyback. Furthermore, the uniform compute design of these batches ameliorates the imbalance between micro-batches, significantly reducing pipeline bubbles. Our techniques yield significant improvements in inference performance across models and hardware. For the LLaMA-13B model on A6000 GPU, SARATHI improves decode throughput by up to 10x, and accelerates end-to-end throughput by up to 1.33x. For LLaMa-33B on A100 GPU, we achieve 1.25x higher end-to-end-throughput and up to 4.25x higher decode throughput. When used with pipeline parallelism on GPT-3, SARATHI reduces bubbles by 6.29x, resulting in an end-to-end throughput improvement of 1.91x.

Anish Biswas, Kanishk Goel, Srivarshinee S et al.

Agentic applications are LLMs that iteratively invoke external tools to accomplish complex tasks. Such tool-based agents are rapidly becoming the dominant paradigm for deploying language models in production. Unlike traditional single-turn inference, agentic workloads chain together multiple LLM calls and tool executions before producing a final response, creating a new performance bottleneck that manifests as increased latency in First Token Rendered (FTR) of the final answer. Through analysis of requests at production scale, we reveal three critical challenges: tool calls account for 30-85% of FTR latency, KV cache hit rates collapse despite substantial context reuse across iterations, and sequential orchestration wastes potential intra-request parallelism. These bottlenecks stem from a design gap in which orchestrators and LLM engines operate as decoupled black boxes, preventing cross-layer optimizations. We present Sutradhara, a co-designed agentic inference system that integrates orchestration with LLM serving through a thin API enabling three optimizations: overlap tool execution with subsequent LLM prefill using tool-aware prompt splitting, streaming tool execution to dispatch tools incrementally during decode rather than waiting for complete output, and orchestrator-aware cache management that uses semantic hints to improve hit rates and reduce thrashing. Implemented on vLLM, Sutradhara improves the throughput-latency trade-off in agentic systems, sustains up to 77% higher load at the same median FTR latency, or reduces median FTR latency by up to 15% at the same load while reducing end-to-end latency by upto 11% on A100 GPUs.

Amey Agrawal, Nitin Kedia, Ashish Panwar et al. · gatech

Each LLM serving request goes through two phases. The first is prefill which processes the entire input prompt and produces the first output token and the second is decode which generates the rest of output tokens, one-at-a-time. Prefill iterations have high latency but saturate GPU compute due to parallel processing of the input prompt. In contrast, decode iterations have low latency but also low compute utilization because a decode iteration processes only a single token per request. This makes batching highly effective for decodes and consequently for overall throughput. However, batching multiple requests leads to an interleaving of prefill and decode iterations which makes it challenging to achieve both high throughput and low latency. We introduce an efficient LLM inference scheduler, Sarathi-Serve, to address this throughput-latency tradeoff. Sarathi-Serve introduces chunked-prefills which splits a prefill request into near equal sized chunks and creates stall-free schedules that adds new requests in a batch without pausing ongoing decodes. Stall-free scheduling unlocks the opportunity to improve throughput with large batch sizes while minimizing the effect of batching on latency. Furthermore, uniform batches in Sarathi-Serve ameliorate the imbalance between iterations resulting in minimal pipeline bubbles. Our techniques yield significant improvements in inference performance across models and hardware under tail latency constraints. For Mistral-7B on single A100 GPUs, we achieve 2.6x higher serving capacity and up to 3.7x higher serving capacity for the Yi-34B model on two A100 GPUs as compared to vLLM. When used with pipeline parallelism on Falcon-180B, Sarathi-Serve provides up to 5.6x gain in the end-to-end serving capacity. The source code for Sarathi-Serve is available at https://github.com/microsoft/sarathi-serve.

Sonu Mehta, Jayashree Mohan, Nagarajan Natarajan et al.

`Extreme Classification'' (or XC) is the task of annotating data points (queries) with relevant labels (documents), from an extremely large set of $L$ possible labels, arising in search and recommendations. The most successful deep learning paradigm that has emerged over the last decade or so for XC is to embed the queries (and labels) using a deep encoder (e.g. DistilBERT), and use linear classifiers on top of the query embeddings. This architecture is of appeal because it enables millisecond-time inference using approximate nearest neighbor search (ANNS). The key question is how do we design training algorithms that are accurate as well as scale to $O(100M)$ labels on a limited number of GPUs. State-of-the-art XC techniques that demonstrate high accuracies (e.g., DEXML, Renée, DEXA) on standard datasets have per-epoch training time that scales as $O(L)$ or employ expensive negative sampling strategies, which are prohibitive in XC scenarios. In this work, we develop an accurate and scalable XC algorithm ASTRA with two key observations: (a) building ANNS index on the classifier vectors and retrieving hard negatives using the classifiers aligns the negative sampling strategy to the loss function optimized; (b) keeping the ANNS indices current as the classifiers change through the epochs is prohibitively expensive while using stale negatives (refreshed periodically) results in poor accuracy; to remedy this, we propose a negative sampling strategy that uses a mixture of importance sampling and uniform sampling. By extensive evaluation on standard XC as well as proprietary datasets with 120M labels, we demonstrate that ASTRA achieves SOTA precision, while reducing training time by 4x-15x relative to the second best.

Amey Agrawal, Nitin Kedia, Jayashree Mohan et al. · gatech

Optimizing the deployment of Large language models (LLMs) is expensive today since it requires experimentally running an application workload against an LLM implementation while exploring large configuration space formed by system knobs such as parallelization strategies, batching techniques, and scheduling policies. To address this challenge, we present Vidur - a large-scale, high-fidelity, easily-extensible simulation framework for LLM inference performance. Vidur models the performance of LLM operators using a combination of experimental profiling and predictive modeling, and evaluates the end-to-end inference performance for different workloads by estimating several metrics of interest such as latency and throughput. We validate the fidelity of Vidur on several LLMs and show that it estimates inference latency with less than 9% error across the range. Further, we present Vidur-Search, a configuration search tool that helps optimize LLM deployment. Vidur-Search uses Vidur to automatically identify the most cost-effective deployment configuration that meets application performance constraints. For example, Vidur-Search finds the best deployment configuration for LLaMA2-70B in one hour on a CPU machine, in contrast to a deployment-based exploration which would require 42K GPU hours - costing ~218K dollars. Source code for Vidur is available at https://github.com/microsoft/vidur.

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.

Aashaka Shah, Chao-Yuan Wu, Jayashree Mohan et al.

Deep learning is slowly, but steadily, hitting a memory bottleneck. While the tensor computation in top-of-the-line GPUs increased by 32x over the last five years, the total available memory only grew by 2.5x. This prevents researchers from exploring larger architectures, as training large networks requires more memory for storing intermediate outputs. In this paper, we present MONeT, an automatic framework that minimizes both the memory footprint and computational overhead of deep networks. MONeT jointly optimizes the checkpointing schedule and the implementation of various operators. MONeT is able to outperform all prior hand-tuned operations as well as automated checkpointing. MONeT reduces the overall memory requirement by 3x for various PyTorch models, with a 9-16% overhead in computation. For the same computation cost, MONeT requires 1.2-1.8x less memory than current state-of-the-art automated checkpointing frameworks. Our code is available at https://github.com/utsaslab/MONeT.

Ramya Prabhu, Ajay Nayak, Jayashree Mohan et al.

PagedAttention is a popular approach for dynamic memory allocation in LLM serving systems. It enables on-demand allocation of GPU memory to mitigate KV cache fragmentation -- a phenomenon that crippled the batch size (and consequently throughput) in prior systems. However, in trying to allocate physical memory at runtime, PagedAttention ends up changing the virtual memory layout of the KV cache from contiguous to non-contiguous. Such a design leads to non-trivial programming and performance overheads. We present vAttention -- an approach that mitigates fragmentation in physical memory while retaining the contiguity of KV cache in virtual memory. We achieve this by decoupling the allocation of virtual and physical memory using CUDA virtual memory management APIs. We also introduce various LLM-specific optimizations to address the limitations of CUDA virtual memory support. Overall, vAttention is a simpler, portable, and performant alternative to PagedAttention: it supports various attention kernels out-of-the-box and improves LLM serving throughput by up to 1.23x compared to the use of PagedAttention-based kernels of FlashAttention and FlashInfer.

Aditya K Kamath, Ramya Prabhu, Jayashree Mohan et al.

Each request in LLM inference goes through two phases: compute-bound prefill and memory-bandwidth-bound decode. To improve GPU utilization, recent systems use hybrid batching that combines the prefill and decode phases of different requests into the same batch. This approach optimizes linear operations but remains inefficient for attention computation because existing attention kernels specialize execution independently for the prefill and decode phases. In this paper, we present POD-Attention - the first GPU kernel that efficiently computes attention for hybrid batches. POD-Attention aims to maximize the utilization of both compute and memory bandwidth by carefully allocating the GPU's resources such that prefill and decode operations happen concurrently on the same multiprocessor. POD-Attention speeds up attention computation by up to $59\%$ (mean $28\%$), enabling higher throughput and lower latency LLM inference compared to the use of independently optimized prefill and decode attention kernels.

Kanishk Goel, Jayashree Mohan, Nipun Kwatra et al.

The widespread adoption of Large Language Models (LLMs) has enabled diverse applications with very different latency requirements. Existing LLM serving frameworks rely on siloed infrastructure with coarse-grained workload segregation -- interactive and batch -- leading to inefficient resource utilization and limited support for fine-grained Quality-of-Service (QoS) differentiation. This results in operational inefficiencies, over-provisioning and poor load management during traffic surges. We present Niyama, a novel QoS-driven inference serving system that enables efficient co-scheduling of diverse workloads on shared infrastructure. Niyama introduces fine-grained QoS classification allowing applications to specify precise latency requirements, and dynamically adapts scheduling decisions based on real-time system state. Leveraging the predictable execution characteristics of LLM inference, Niyama implements a dynamic chunking mechanism to improve overall throughput while maintaining strict QoS guarantees. Additionally, Niyama employs a hybrid prioritization policy that balances fairness and efficiency, and employs selective request relegation that enables graceful service degradation during overload conditions. Our evaluation demonstrates that Niyama increases serving capacity by 32% compared to current siloed deployments, while maintaining QoS guarantees. Notably, under extreme load, our system reduces SLO violations by an order of magnitude compared to current strategies.

Amey Agrawal, Nitin Kedia, Anmol Agarwal et al. · gatech

The rapid evolution of Large Language Model (LLM) inference systems has yielded significant efficiency improvements. However, our systematic analysis reveals that current evaluation methodologies frequently exhibit fundamental flaws, often manifesting as common evaluation anti-patterns that obscure true performance characteristics and impede scientific progress. Through a comprehensive examination of recent systems, we identify recurring anti-patterns across three key dimensions: Baseline Fairness, Evaluation Setup, and Metric Design. These anti-patterns are uniquely problematic for LLM inference due to its dual-phase nature combining distinct prefill and decode operations, its handling of highly heterogeneous workloads, and its strict temporal requirements for interactive use. We demonstrate how common anti-patterns -- such as inadequate baseline comparisons that conflate engineering effort with algorithmic novelty, workload selections that fail to represent production scenarios, and metric normalizations that hide substantial performance variability like generation stalls-lead to misleading conclusions. To address these challenges, we provide a comprehensive checklist derived from our analysis, establishing a framework for recognizing and avoiding these anti-patterns in favor of robust LLM inference evaluation. To demonstrate the practical application of our framework, we present a case study analyzing speculative decoding, a technique whose bursty, non-uniform token generation is easily misinterpreted when evaluated using approaches characteristic of these anti-patterns. Our work establishes a rigorous foundation for evaluation methodology, enabling meaningful comparisons, ensuring reproducible results, and ultimately accelerating genuine progress in LLM inference systems by moving beyond common anti-patterns to align evaluation with real-world requirements.

Tella Rajashekhar Reddy, Palak, Rohan Gandhi et al.

AI power demand is growing unprecedentedly thanks to the high power density of AI compute and the emerging inferencing workload. On the supply side, abundant wind power is waiting for grid access in interconnection queues. In this light, this paper argues bringing AI workload to modular compute clusters co-located in wind farms. Our deployment right-sizing strategy makes it economically viable to deploy more than 6 million high-end GPUs today that could consume cheap, green power at its source. We built Heron, a cross-site software router, that could efficiently leverage the complementarity of power generation across wind farms by routing AI inferencing workload around power drops. Using 1-week ofcoding and conversation production traces from Azure and (real) variable wind power traces, we show how Heron improves aggregate goodput of AI compute by up to 80% compared to the state-of-the-art.

Jayashree Mohan, Amar Phanishayee, Janardhan Kulkarni et al.

Training Deep Neural Networks (DNNs) is a widely popular workload in both enterprises and cloud data centers. Existing schedulers for DNN training consider GPU as the dominant resource, and allocate other resources such as CPU and memory proportional to the number of GPUs requested by the job. Unfortunately, these schedulers do not consider the impact of a job's sensitivity to allocation of CPU, memory, and storage resources. In this work, we propose Synergy, a resource-sensitive scheduler for shared GPU clusters. Synergy infers the sensitivity of DNNs to different resources using optimistic profiling; some jobs might benefit from more than the GPU-proportional allocation and some jobs might not be affected by less than GPU-proportional allocation. Synergy performs such multi-resource workload-aware assignments across a set of jobs scheduled on shared multi-tenant clusters using a new near-optimal online algorithm. Our experiments show that workload-aware CPU and memory allocations can improve average JCT up to 3.4x when compared to traditional GPU-proportional scheduling.

Jayashree Mohan, Amar Phanishayee, Ashish Raniwala et al.

Training Deep Neural Networks (DNNs) is resource-intensive and time-consuming. While prior research has explored many different ways of reducing DNN training time, the impact of input data pipeline, i.e., fetching raw data items from storage and performing data pre-processing in memory, has been relatively unexplored. This paper makes the following contributions: (1) We present the first comprehensive analysis of how the input data pipeline affects the training time of widely-used computer vision and audio Deep Neural Networks (DNNs), that typically involve complex data preprocessing. We analyze nine different models across three tasks and four datasets while varying factors such as the amount of memory, number of CPU threads, storage device, GPU generation etc on servers that are a part of a large production cluster at Microsoft. We find that in many cases, DNN training time is dominated by data stall time: time spent waiting for data to be fetched and preprocessed. (2) We build a tool, DS-Analyzer to precisely measure data stalls using a differential technique, and perform predictive what-if analysis on data stalls. (3) Finally, based on the insights from our analysis, we design and implement three simple but effective techniques in a data-loading library, CoorDL, to mitigate data stalls. Our experiments on a range of DNN tasks, models, datasets, and hardware configs show that when PyTorch uses CoorDL instead of the state-of-the-art DALI data loading library, DNN training time is reduced significantly (by as much as 5x on a single server).