Aditya Akella

Sudarsanan Rajasekaran, Manya Ghobadi, Aditya Akella

We present CASSINI, a network-aware job scheduler for machine learning (ML) clusters. CASSINI introduces a novel geometric abstraction to consider the communication pattern of different jobs while placing them on network links. To do so, CASSINI uses an affinity graph that finds a series of time-shift values to adjust the communication phases of a subset of jobs, such that the communication patterns of jobs sharing the same network link are interleaved with each other. Experiments with 13 common ML models on a 24-server testbed demonstrate that compared to the state-of-the-art ML schedulers, CASSINI improves the average and tail completion time of jobs by up to 1.6x and 2.5x, respectively. Moreover, we show that CASSINI reduces the number of ECN marked packets in the cluster by up to 33x.

Jiachen Liu, Fan Lai, Yinwei Dai et al.

Federated learning (FL) is an emerging machine learning (ML) paradigm that enables heterogeneous edge devices to collaboratively train ML models without revealing their raw data to a logically centralized server. However, beyond the heterogeneous device capacity, FL participants often exhibit differences in their data distributions, which are not independent and identically distributed (Non-IID). Many existing works present point solutions to address issues like slow convergence, low final accuracy, and bias in FL, all stemming from client heterogeneity. In this paper, we explore an additional layer of complexity to mitigate such heterogeneity by grouping clients with statistically similar data distributions (cohorts). We propose Auxo to gradually identify such cohorts in large-scale, low-availability, and resource-constrained FL populations. Auxo then adaptively determines how to train cohort-specific models in order to achieve better model performance and ensure resource efficiency. Our extensive evaluations show that, by identifying cohorts with smaller heterogeneity and performing efficient cohort-based training, Auxo boosts various existing FL solutions in terms of final accuracy (2.1% - 8.2%), convergence time (up to 2.2x), and model bias (4.8% - 53.8%).

Tarannum Khan, Saeed Rashidi, Srinivas Sridharan et al.

RDMA over Converged Ethernet (RoCE) has gained significant attraction for datacenter networks due to its compatibility with conventional Ethernet-based fabric. However, the RDMA protocol is efficient only on (nearly) lossless networks, emphasizing the vital role of congestion control on RoCE networks. Unfortunately, the native RoCE congestion control scheme, based on Priority Flow Control (PFC), suffers from many drawbacks such as unfairness, head-of-line-blocking, and deadlock. Therefore, in recent years many schemes have been proposed to provide additional congestion control for RoCE networks to minimize PFC drawbacks. However, these schemes are proposed for general datacenter environments. In contrast to the general datacenters that are built using commodity hardware and run general-purpose workloads, high-performance distributed training platforms deploy high-end accelerators and network components and exclusively run training workloads using collectives (All-Reduce, All-To-All) communication libraries for communication. Furthermore, these platforms usually have a private network, separating their communication traffic from the rest of the datacenter traffic. Scalable topology-aware collective algorithms are inherently designed to avoid incast patterns and balance traffic optimally. These distinct features necessitate revisiting previously proposed congestion control schemes for general-purpose datacenter environments. In this paper, we thoroughly analyze some of the SOTA RoCE congestion control schemes vs. PFC when running on distributed training platforms. Our results indicate that previously proposed RoCE congestion control schemes have little impact on the end-to-end performance of training workloads, motivating the necessity of designing an optimized, yet low-overhead, congestion control scheme based on the characteristics of distributed training platforms and workloads.

Robert Grandl, Srikanth Kandula, Sriram Rao et al.

We present a scheduler that improves cluster utilization and job completion times by packing tasks having multi-resource requirements and inter-dependencies. While the problem is algorithmically very hard, we achieve near-optimality on the job DAGs that appear in production clusters at a large enterprise and in benchmarks such as TPC-DS. A key insight is that carefully handling the long-running tasks and those with tough-to-pack resource needs will produce good-enough schedules. However, which subset of tasks to treat carefully is not clear (and intractable to discover). Hence, we offer a search procedure that evaluates various possibilities and outputs a preferred schedule order over tasks. An online component enforces the schedule orders desired by the various jobs running on the cluster. In addition, it packs tasks, overbooks the fungible resources and guarantees bounded unfairness for a variety of desirable fairness schemes. Relative to the state-of-the art schedulers, we speed up 50% of the jobs by over 30% each.

Jiajun Zhu, Ying Chen, Peihao Wang et al.

Graph foundation models (GFMs) aim to reuse a single backbone across diverse graph domains, yet their transfer is often uneven and can exhibit negative transfer. While most prior work improves transfer through architectural or adaptation choices, we ask a data-centric question: which properties of two graph domains determine how much a fixed representation model changes its outputs? Using a graphon-based continuous limit for dense graphs, we show that for both set-based and message-passing tokenizations, any Lipschitz backbone admits an explicit decomposition of cross-domain output shift into (i) graph-specific finite-sample approximation terms and (ii) an intrinsic, relabeling-invariant domain discrepancy capturing structural mismatch. A key ingredient is positional-encoding (PE) stability: we establish stability guarantees for spectral PEs and highlight contrasting behaviors of eigenvector- versus subspace-based PEs. Experiments on synthetic and real graphs validate the theory and translate the decomposition into guidance for data curation in GFM transfer.

Jianing Zhu, Yeonju Ro, John Robertson et al.

Long-lived AI agents are increasingly deployed as persistent operational systems, yet they are still evaluated like freshly initialized models. Day-one benchmarks miss a basic systems question: how long does an agent remain reliable after deployment? Even when model weights are frozen, an agent's effective state keeps changing as it compresses interaction history, retrieves from a growing memory store, revises facts after updates, and undergoes routine maintenance. Reliability therefore becomes a lifespan property of the full agent harness, not only a snapshot property of the base model. We introduce AgingBench, a longitudinal reliability benchmark for agent lifespan engineering: measuring not only whether deployed agents degrade, but what form the degradation takes and where repair should target. AgingBench organizes agent aging into four mechanisms: compression aging, interference aging, revision aging, and maintenance aging. To diagnose these failures, AgingBench uses temporal dependency graphs and paired counterfactual probes that produce diagnostic profiles for the write, retrieval, and utilization stages of the memory pipeline. Across 7 scenarios, 14 models, multiple memory policies, and both runner-controlled and autonomous agents, over ~400 runs spanning 8 - 200 sessions show that agent aging is not one-dimensional: behavioral tests can remain clean while factual precision decays; derived-state tracking can collapse sharply within a single model; and the same wrong answer can require different repairs depending on what the diagnostic profile points to. These results suggest that reliable agent deployment requires lifespan evaluation, mechanism-level diagnosis, and stage-targeted repair, not only stronger day-one models.

Bodun Hu, Le Xu, Jeongyoon Moon et al.

Rapid advancements over the years have helped machine learning models reach previously hard-to-achieve goals, sometimes even exceeding human capabilities. However, to attain the desired accuracy, the model sizes and in turn their computational requirements have increased drastically. Thus, serving predictions from these models to meet any target latency and cost requirements of applications remains a key challenge, despite recent work in building inference-serving systems as well as algorithmic approaches that dynamically adapt models based on inputs. In this paper, we introduce a form of dynamism, modality selection, where we adaptively choose modalities from inference inputs while maintaining the model quality. We introduce MOSEL, an automated inference serving system for multi-modal ML models that carefully picks input modalities per request based on user-defined performance and accuracy requirements. MOSEL exploits modality configurations extensively, improving system throughput by 3.6$\times$ with an accuracy guarantee and shortening job completion times by 11$\times$.

Chi Zhang, Olga Papaemmanouil, Josiah P. Hanna et al.

Machine learning is rapidly being used in database research to improve the effectiveness of numerous tasks included but not limited to query optimization, workload scheduling, physical design, etc. Currently, the research focus has been on replacing a single database component responsible for one task by its learning-based counterpart. However, query performance is not simply determined by the performance of a single component, but by the cooperation of multiple ones. As such, learning based database components need to collaborate during both training and execution in order to develop policies that meet end performance goals. Thus, the paper attempts to address the question "Is it possible to design a database consisting of various learned components that cooperatively work to improve end-to-end query latency?". To answer this question, we introduce MADB (Multi-Agent DB), a proof-of-concept system that incorporates a learned query scheduler and a learned query optimizer. MADB leverages a cooperative multi-agent reinforcement learning approach that allows the two components to exchange the context of their decisions with each other and collaboratively work towards reducing the query latency. Preliminary results demonstrate that MADB can outperform the non-cooperative integration of learned components.

Ruisi Cai, Yeonju Ro, Geon-Woo Kim et al.

The proliferation of large language models (LLMs) has led to the adoption of Mixture-of-Experts (MoE) architectures that dynamically leverage specialized subnetworks for improved efficiency and performance. Despite their benefits, MoE models face significant challenges during inference, including inefficient memory management and suboptimal batching, due to misaligned design choices between the model architecture and the system policies. Furthermore, the conventional approach of training MoEs from scratch is increasingly prohibitive in terms of cost. In this paper, we propose a novel framework Read-ME that transforms pre-trained dense LLMs into smaller MoE models (in contrast to "upcycling" generalist MoEs), avoiding the high costs of ground-up training. Our approach employs activation sparsity to extract experts. To compose experts, we examine the widely-adopted layer-wise router design and show its redundancy, and thus we introduce the pre-gating router decoupled from the MoE backbone that facilitates system-friendly pre-computing and lookahead scheduling, enhancing expert-aware batching and caching. Our codesign therefore addresses critical gaps on both the algorithmic and system fronts, establishing a scalable and efficient alternative for LLM inference in resource-constrained settings. Read-ME outperforms other popular open-source dense models of similar scales, achieving improvements of up to 10.1% on MMLU, and improving mean end-to-end latency up to 6.1%. Codes are available at: https://github.com/VITA-Group/READ-ME.

Yeonju Ro, Zhenyu Zhang, Souvik Kundu et al.

Large language models (LLMs) excel at capturing global token dependencies via self-attention but face prohibitive compute and memory costs on lengthy inputs. While sub-quadratic methods (e.g., linear attention) can reduce these costs, they often degrade accuracy due to overemphasizing recent tokens. In this work, we first propose dual-state linear attention (DSLA), a novel design that maintains two specialized hidden states-one for preserving historical context and one for tracking recency-thereby mitigating the short-range bias typical of linear-attention architectures. To further balance efficiency and accuracy under dynamic workload conditions, we introduce DSLA-Serve, an online adaptive distillation framework that progressively replaces Transformer layers with DSLA layers at inference time, guided by a sensitivity-based layer ordering. DSLA-Serve uses a chained fine-tuning strategy to ensure that each newly converted DSLA layer remains consistent with previously replaced layers, preserving the overall quality. Extensive evaluations on commonsense reasoning, long-context QA, and text summarization demonstrate that DSLA-Serve yields 2.3x faster inference than Llama2-7B and 3.0x faster than the hybrid Zamba-7B, while retaining comparable performance across downstream tasks. Our ablation studies show that DSLA's dual states capture both global and local dependencies, addressing the historical-token underrepresentation seen in prior linear attentions. Codes are available at https://github.com/utnslab/DSLA-Serve.

Rohit Dwivedula, Divyanshu Saxena, Aditya Akella et al.

Policy design for various systems controllers has conventionally been a manual process, with domain experts carefully tailoring heuristics for the specific instance in which the policy will be deployed. In this paper, we re-imagine policy design via a novel automated search technique fueled by recent advances in generative models, specifically Large Language Model (LLM)-driven code generation. We outline the design and implementation of PolicySmith, a framework that applies LLMs to synthesize instance-optimal heuristics. We apply PolicySmith to two long-standing systems policies - web caching and congestion control, highlighting the opportunities unraveled by this LLM-driven heuristic search. For caching, PolicySmith discovers heuristics that outperform established baselines on standard open-source traces. For congestion control, we show that PolicySmith can generate safe policies that integrate directly into the Linux kernel.

Jucheng Shen, Gaurav Sarkar, Yeonju Ro et al.

We present CadLLM, a training-free method to accelerate the inference throughput of diffusion-based LLMs (dLLMs). We first investigate the dynamic nature of token unmasking confidence across blocks and steps. Based on this observation, we present a lightweight adaptive approach that controls the generation block size, step size, and threshold based on the average confidence of unmasked tokens. We further reduce softmax overhead by dynamically leveraging a subset of the vocabulary to regulate sampling breadth. CadLLM is a plug-and-play, model-agnostic method compatible with KV-cache-based dLLMs. Extensive experiments on four popular tasks demonstrate that CadLLM yields up to 2.28x throughput improvement over the state-of-the-art baseline with competitive accuracy.

Jimmy Shong, Yuhan Ding, Yihan Jiang et al.

Machine learning (ML) systems expose a rapidly expanding configuration space spanning model-parallelism strategies, communication optimizations, and low-level runtime parameters. End-to-end system efficiency is highly sensitive to these choices, yet identifying high-performance configurations is challenging due to heterogeneous feature types (e.g., sparse and dense parameters), conditional dependencies (e.g., valid execution parameters only under specific upstream decisions), and the high search (profiling) cost. Existing approaches either optimize a narrow subset of configuration dimensions or rely on ad-hoc heuristics that fail to generalize as configuration spaces continue to grow. We present AutoScout, a general-purpose systems configurator for ML training, fine-tuning, and inference. It formulates the system configuration as a mixed-discrete/continuous optimization problem with hierarchical dependencies and introduces a hybrid optimization framework that jointly refines sparse structural decisions and dense execution parameters. To reduce profiling cost, AutoScout adaptively prioritizes high-impact configuration features and ensembles simulators with varying fidelity. Across diverse models, hardware platforms, and deployment objectives, AutoScout consistently identifies high-performance configurations, achieving 2.7-3.0$\times$ training speedup over expert-tuned settings.

Rohit Dwivedula, Divyanshu Saxena, Sujay Yadalam et al.

Resource-management tasks in modern operating and distributed systems continue to rely primarily on hand-designed heuristics for tasks such as scheduling, caching, or active queue management. Designing performant heuristics is an expensive, time-consuming process that we are forced to continuously go through due to the constant flux of hardware, workloads and environments. We propose a new alternative: synthesizing instance-optimal heuristics -- specialized for the exact workloads and hardware where they will be deployed -- using code-generating large language models (LLMs). To make this synthesis tractable, Vulcan separates policy and mechanism through LLM-friendly, task-agnostic interfaces. With these interfaces, users specify the inputs and objectives of their desired policy, while Vulcan searches for performant policies via evolutionary search over LLM-generated code. This interface is expressive enough to capture a wide range of system policies, yet sufficiently constrained to allow even small, inexpensive LLMs to generate correct and executable code. We use Vulcan to synthesize performant heuristics for cache eviction and memory tiering, and find that these heuristics outperform all human-designed state-of-the-art algorithms by upto 69% and 7.9% in performance for each of these tasks respectively.

Jeremy Carleton, Prathik Vijaykumar, Divyanshu Saxena et al.

We address the challenge of zeroth-order online convex optimization where the objective function's gradient exhibits sparsity, indicating that only a small number of dimensions possess non-zero gradients. Our aim is to leverage this sparsity to obtain useful estimates of the objective function's gradient even when the only information available is a limited number of function samples. Our motivation stems from the optimization of large-scale queueing networks that process time-sensitive jobs. Here, a job must be processed by potentially many queues in sequence to produce an output, and the service time at any queue is a function of the resources allocated to that queue. Since resources are costly, the end-to-end latency for jobs must be balanced with the overall cost of the resources used. While the number of queues is substantial, the latency function primarily reacts to resource changes in only a few, rendering the gradient sparse. We tackle this problem by introducing the Compressive Online Gradient Optimization framework which allows compressive sensing methods previously applied to stochastic optimization to achieve regret bounds with an optimal dependence on the time horizon without the full problem dimension appearing in the bound. For specific algorithms, we reduce the samples required per gradient estimate to scale with the gradient's sparsity factor rather than its full dimensionality. Numerical simulations and real-world microservices benchmarks demonstrate CONGO's superiority over gradient descent approaches that do not account for sparsity.

Bodun Hu, Yoga Sri Varshan, Saurabh Agarwal et al.

Custom CUDA kernel development is essential for maximizing GPU utilization in large-scale distributed LLM training and inference, yet manually writing kernels that jointly leverage both computation and communication remains a labor-intensive and error-prone process. Prior work on kernel optimization has focused almost exclusively on computation, leaving communication kernels largely untouched even though they constitute a significant share of total execution time. We introduce CUCo, a training-free agent-driven workflow that automatically generates high-performance CUDA kernels that jointly orchestrate computation and communication. By co-optimizing these traditionally disjoint components, CUCo unlocks new optimization opportunities unavailable to existing approaches, outperforming state-of-the-art baselines and reducing end-to-end latency by up to $1.57\times$.

Ajay Jaiswal, Bodun Hu, Lu Yin et al.

Autoregressive Large Language Models (e.g., LLaMa, GPTs) are omnipresent achieving remarkable success in language understanding and generation. However, such impressive capability typically comes with a substantial model size, which presents significant challenges for autoregressive token-by-token generation. To mitigate computation overload incurred during generation, several early-exit and layer-dropping strategies have been proposed. Despite some promising success due to the redundancy across LLMs layers on metrics like Rough-L/BLUE, our careful knowledge-intensive evaluation unveils issues such as generation collapse, hallucination of wrong facts, and noticeable performance drop even at the trivial exit ratio of 10-15% of layers. We attribute these errors primarily to ineffective handling of the KV cache through state copying during early-exit. In this work, we observed the saturation of computationally expensive feed-forward blocks of LLM layers and proposed FFN-SkipLLM, which is a novel fine-grained skip strategy of autoregressive LLMs. More specifically, FFN-SkipLLM is an input-adaptive feed-forward skipping strategy that can skip 25-30% of FFN blocks of LLMs with marginal change in performance on knowledge-intensive generation tasks without any requirement to handle KV cache. Our extensive experiments and ablation across benchmarks like MT-Bench, Factoid-QA, and variable-length text summarization illustrate how our simple and ease-at-use method can facilitate faster autoregressive decoding.

Haoyu Li, Yuchen Xu, Jiayi Chen et al.

As deep neural networks (DNNs) grow in complexity and size, the resultant increase in communication overhead during distributed training has become a significant bottleneck, challenging the scalability of distributed training systems. Existing solutions, while aiming to mitigate this bottleneck through worker-level compression and in-network aggregation, fall short due to their inability to efficiently reconcile the trade-offs between compression effectiveness and computational overhead, hindering overall performance and scalability. In this paper, we introduce a novel compression algorithm that effectively merges worker-level compression with in-network aggregation. Our solution is both homomorphic, allowing for efficient in-network aggregation without CPU/GPU processing, and lossless, ensuring no compromise on training accuracy. Theoretically optimal in compression and computational efficiency, our approach is empirically validated across diverse DNN models such as NCF, LSTM, VGG19, and BERT-base, showing up to a 6.33$\times$ improvement in aggregation throughput and a 3.74$\times$ increase in per-iteration training speed.

Geon-Woo Kim, Donghyun Kim, Jeongyoon Moon et al.

Graph Neural Networks (GNNs) have been widely adopted for their ability to compute expressive node representations in graph datasets. However, serving GNNs on large graphs is challenging due to the high communication, computation, and memory overheads of constructing and executing computation graphs, which represent information flow across large neighborhoods. Existing approximation techniques in training can mitigate the overheads but, in serving, still lead to high latency and/or accuracy loss. To this end, we propose OMEGA, a system that enables low-latency GNN serving for large graphs with minimal accuracy loss through two key ideas. First, OMEGA employs selective recomputation of precomputed embeddings, which allows for reusing precomputed computation subgraphs while selectively recomputing a small fraction to minimize accuracy loss. Second, we develop computation graph parallelism, which reduces communication overhead by parallelizing the creation and execution of computation graphs across machines. Our evaluation with large graph datasets and GNN models shows that OMEGA significantly outperforms state-of-the-art techniques.

Divyanshu Saxena, Nihal Sharma, Donghyun Kim et al.

This paper lays down the research agenda for a domain-specific foundation model for operating systems (OSes). Our case for a foundation model revolves around the observations that several OS components such as CPU, memory, and network subsystems are interrelated and that OS traces offer the ideal dataset for a foundation model to grasp the intricacies of diverse OS components and their behavior in varying environments and workloads. We discuss a wide range of possibilities that then arise, from employing foundation models as policy agents to utilizing them as generators and predictors to assist traditional OS control algorithms. Our hope is that this paper spurs further research into OS foundation models and creating the next generation of operating systems for the evolving computing landscape.

Donghyun Kim, Sriram Ravula, Taemin Ha et al. · microsoft-research

Workload traces are essential to understand complex computer systems' behavior and manage processing and memory resources. Since real-world traces are hard to obtain, synthetic trace generation is a promising alternative. This paper proposes a first-of-a-kind approach that relies on training a large language model (LLM) to generate synthetic workload traces, specifically microservice call graphs. To capture complex and arbitrary hierarchical structures and implicit constraints in such traces, we propose to train LLMs to generate recursively, making call graph generation a sequence of more manageable steps. To further enforce learning constraints on the traces and generate uncommon situations, we apply additional instruction tuning steps to align our model with the desired trace features. With this method, we train TraceLLM, an LLM for microservice trace generation, and demonstrate that it produces diverse, realistic traces under varied conditions, outperforming existing approaches in both accuracy and validity. The synthetically generated traces can effectively replace real data to optimize important microservice management tasks. Additionally, TraceLLM adapts to downstream trace-related tasks, such as predicting key trace features and infilling missing data.

Geon-Woo Kim, Junbo Li, Shashidhar Gandham et al.

Training large language models (LLMs) increasingly relies on geographically distributed accelerators, causing prohibitive communication costs across regions and uneven utilization of heterogeneous hardware. We propose HALoS, a hierarchical asynchronous optimization framework that tackles these issues by introducing local parameter servers (LPSs) within each region and a global parameter server (GPS) that merges updates across regions. This hierarchical design minimizes expensive inter-region communication, reduces straggler effects, and leverages fast intra-region links. We provide a rigorous convergence analysis for HALoS under non-convex objectives, including theoretical guarantees on the role of hierarchical momentum in asynchronous training. Empirically, HALoS attains up to 7.5x faster convergence than synchronous baselines in geo-distributed LLM training and improves upon existing asynchronous methods by up to 2.1x. Crucially, HALoS preserves the model quality of fully synchronous SGD-matching or exceeding accuracy on standard language modeling and downstream benchmarks-while substantially lowering total training time. These results demonstrate that hierarchical, server-side update accumulation and global model merging are powerful tools for scalable, efficient training of new-era LLMs in heterogeneous, geo-distributed environments.

ChonLam Lao, Minlan Yu, Aditya Akella et al.

Large-scale ML training jobs are frequently interrupted by hardware and software anomalies, failures, and management events. Existing solutions like checkpointing or runtime reconfiguration suffer from long downtimes, degraded performance, or undesired changes to training strategies. We present TrainMover, a resilient runtime that leverages standby machines to handle interruptions with minimal downtime and zero memory overhead. To achieve these goals, TrainMover introduces two key techniques: two-phase, delta-based communication group setups and communication-free sandboxed shadow iterations. Our evaluation shows that TrainMover consistently achieves second-level downtime across all evaluated models during migration, maintaining 99\% training efficiency during periodic 10-minute rebalancing. We also demonstrate the effectiveness of TrainMover in handling various interruptions.

Chenxi Yang, Divyanshu Saxena, Rohit Dwivedula et al.

Learning-based congestion controllers offer better adaptability compared to traditional heuristic algorithms. However, the inherent unreliability of learning techniques can cause learning-based controllers to behave poorly, creating a need for formal guarantees. While methods for formally verifying learned congestion controllers exist, these methods offer binary feedback that cannot optimize the controller toward better behavior. We improve this state-of-the-art via C3, a new learning framework for congestion control that integrates the concept of formal certification in the learning loop. C3 uses an abstract interpreter that can produce robustness and performance certificates to guide the training process, rewarding models that are robust and performant even on worst-case inputs. Our evaluation demonstrates that unlike state-of-the-art learned controllers, C3-trained controllers provide both adaptability and worst-case reliability across a range of network conditions.

Aditya Akella

Recent advances in learnable reward shaping have shown promise in single-agent reinforcement learning by automatically discovering effective feedback signals. However, the effectiveness of decentralized learnable reward shaping in cooperative multi-agent settings remains poorly understood. We propose DMARL-RSA, a fully decentralized system where each agent learns individual reward shaping, and evaluate it on cooperative navigation tasks in the simple_spread_v3 environment. Despite sophisticated reward learning, DMARL-RSA achieves only -24.20 +/- 0.09 average reward, compared to MAPPO with centralized training at 1.92 +/- 0.87 -- a 26.12-point gap. DMARL-RSA performs similarly to simple independent learning (IPPO: -23.19 +/- 0.96), indicating that advanced reward shaping cannot overcome fundamental decentralized coordination limitations. Interestingly, decentralized methods achieve higher landmark coverage (0.888 +/- 0.029 for DMARL-RSA, 0.960 +/- 0.045 for IPPO out of 3 total) but worse overall performance than centralized MAPPO (0.273 +/- 0.008 landmark coverage) -- revealing a coordination paradox between local optimization and global performance. Analysis identifies three critical barriers: (1) non-stationarity from concurrent policy updates, (2) exponential credit assignment complexity, and (3) misalignment between individual reward optimization and global objectives. These results establish empirical limits for decentralized reward learning and underscore the necessity of centralized coordination for effective multi-agent cooperation.

Samuel Yuan, Divyanshu Saxena, Jiayi Chen et al.

Several learned policies have been proposed to replace heuristics for scheduling, caching, and other system components in modern systems. By leveraging diverse features, learning from historical trends, and predicting future behaviors, such models promise to keep pace with ever-increasing workload dynamism and continuous hardware evolution. However, policies trained in isolation may still achieve suboptimal performance when placed together. In this paper, we inspect one such instance in the domain of hardware caching -- for the policies of cache replacement and prefetching. We argue that these two policies are bidirectionally interdependent and make the case for training the two jointly. We propose a joint learning approach based on developing shared representations for the features used by the two policies. We present two approaches to develop these shared representations, one based on a joint encoder and another based on contrastive learning of the embeddings, and demonstrate promising preliminary results for both of these. Finally, we lay down an agenda for future research in this direction.

Fanjiang Ye, Zepeng Zhao, Yi Mu et al.

Diffusion models have recently achieved remarkable success in generative tasks (e.g., image and video generation), and the demand for high-quality content (e.g., 2K/4K videos) is rapidly increasing across various domains. However, generating ultra-high-resolution videos on existing standard-resolution (e.g., 720p) platforms remains challenging due to the excessive re-training requirements and prohibitively high computational and memory costs. To this end, we introduce SuperGen, an efficient tile-based framework for ultra-high-resolution video generation. SuperGen features a novel training-free algorithmic innovation with tiling to successfully support a wide range of resolutions without additional training efforts while significantly reducing both memory footprint and computational complexity. Moreover, SuperGen incorporates a tile-tailored, adaptive, region-aware caching strategy that accelerates video generation by exploiting redundancy across denoising steps and spatial regions. SuperGen also integrates cache-guided, communication-minimized tile parallelism for enhanced throughput and minimized latency. Evaluations demonstrate that SuperGen harvests the maximum performance gains while achieving high output quality across various benchmarks.

Aditya Akella, Jonathan Farah

Demographic shifts, influenced by globalization, economic conditions, geopolitical events, and environmental factors, pose significant challenges for policymakers and researchers. Accurate demographic forecasting is essential for informed decision-making in areas such as urban planning, healthcare, and economic policy. This study explores the application of time series foundation models to predict demographic changes in the United States using datasets from the U.S. Census Bureau and Federal Reserve Economic Data (FRED). We evaluate the performance of the Time Series Foundation Model (TimesFM) against traditional baselines including Long Short-Term Memory (LSTM) networks, Autoregressive Integrated Moving Average (ARIMA), and Linear Regression. Our experiments across six demographically diverse states demonstrate that TimesFM achieves the lowest Mean Squared Error (MSE) in 86.67% of test cases, with particularly strong performance on minority populations with sparse historical data. These findings highlight the potential of pre-trained foundation models to enhance demographic analysis and inform proactive policy interventions without requiring extensive task-specific fine-tuning.

Bodun Hu, Luis Pabon, Saurabh Agarwal et al.

Retrieval Augmented Generation (RAG) has emerged as a new paradigm for enhancing Large Language Model reliability through integration with external knowledge sources. However, efficient deployment of these systems presents significant technical challenges due to their inherently heterogeneous computational pipelines comprising LLMs, databases, and specialized processing components. We introduce Patchwork, a comprehensive end-to-end RAG serving framework designed to address these efficiency bottlenecks. Patchwork's architecture offers three key innovations: First, it provides a flexible specification interface enabling users to implement custom RAG pipelines. Secondly, it deploys these pipelines as distributed inference systems while optimizing for the unique scalability characteristics of individual RAG components. Third, Patchwork incorporates an online scheduling mechanism that continuously monitors request load and execution progress, dynamically minimizing SLO violations through strategic request prioritization and resource auto-scaling. Our experimental evaluation across four distinct RAG implementations demonstrates that Patchwork delivers substantial performance improvements over commercial alternatives, achieving throughput gains exceeding 48% while simultaneously reducing SLO violations by ~24%.

Haoyu Li, Srikanth Kandula, Maria Angels de Luis Balaguer et al.

Analyzing large datasets requires responsive query execution, but executing SQL queries on massive datasets can be slow. This paper explores whether query execution can begin even before the user has finished typing, allowing results to appear almost instantly. We propose SpeQL, a system that leverages Large Language Models (LLMs) to predict likely queries based on the database schema, the user's past queries, and their incomplete query. Since exact query prediction is infeasible, SpeQL speculates on partial queries in two ways: 1) it predicts the query structure to compile and plan queries in advance, and 2) it precomputes smaller temporary tables that are much smaller than the original database, but are still predicted to contain all information necessary to answer the user's final query. Additionally, SpeQL continuously displays results for speculated queries and subqueries in real time, aiding exploratory analysis. A utility/user study showed that SpeQL improved task completion time, and participants reported that its speculative display of results helped them discover patterns in the data more quickly. In the study, SpeQL improves user's query latency by up to $289\times$ and kept the overhead reasonable, at $\$4$ per hour.

Adarsh Kumar, Kausik Subramanian, Shivaram Venkataraman et al.

Many organizations employ compute clusters equipped with accelerators such as GPUs and TPUs for training deep learning models in a distributed fashion. Training is resource-intensive, consuming significant compute, memory, and network resources. Many prior works explore how to reduce training resource footprint without impacting quality, but their focus on a subset of the bottlenecks (typically only the network) limits their ability to improve overall cluster utilization. In this work, we exploit the unique characteristics of deep learning workloads to propose Structured Partial Backpropagation(SPB), a technique that systematically controls the amount of backpropagation at individual workers in distributed training. This simultaneously reduces network bandwidth, compute utilization, and memory footprint while preserving model quality. To efficiently leverage the benefits of SPB at cluster level, we introduce JigSaw, a SPB aware scheduler, which does scheduling at the iteration level for Deep Learning Training(DLT) jobs. We find that JigSaw can improve large scale cluster efficiency by as high as 28\%.

Arjun Balasubramanian, Adarsh Kumar, Yuhan Liu et al.

Deep Neural Networks (DNNs) are witnessing increased adoption in multiple domains owing to their high accuracy in solving real-world problems. However, this high accuracy has been achieved by building deeper networks, posing a fundamental challenge to the low latency inference desired by user-facing applications. Current low latency solutions trade-off on accuracy or fail to exploit the inherent temporal locality in prediction serving workloads. We observe that caching hidden layer outputs of the DNN can introduce a form of late-binding where inference requests only consume the amount of computation needed. This enables a mechanism for achieving low latencies, coupled with an ability to exploit temporal locality. However, traditional caching approaches incur high memory overheads and lookup latencies, leading us to design learned caches - caches that consist of simple ML models that are continuously updated. We present the design of GATI, an end-to-end prediction serving system that incorporates learned caches for low-latency DNN inference. Results show that GATI can reduce inference latency by up to 7.69X on realistic workloads.

Adarsh Kumar, Arjun Balasubramanian, Shivaram Venkataraman et al.

Over the last few years, Deep Neural Networks (DNNs) have become ubiquitous owing to their high accuracy on real-world tasks. However, this increase in accuracy comes at the cost of computationally expensive models leading to higher prediction latencies. Prior efforts to reduce this latency such as quantization, model distillation, and any-time prediction models typically trade-off accuracy for performance. In this work, we observe that caching intermediate layer outputs can help us avoid running all the layers of a DNN for a sizeable fraction of inference requests. We find that this can potentially reduce the number of effective layers by half for 91.58% of CIFAR-10 requests run on ResNet-18. We present Freeze Inference, a system that introduces approximate caching at each intermediate layer and we discuss techniques to reduce the cache size and improve the cache hit rate. Finally, we discuss some of the open research challenges in realizing such a design.