arXiv subjects

Dmitrii Ustiugov

Publications and source records attributed to Dmitrii Ustiugov.

16 recordsLinked to original sources

arXiv2026

CodecSight: Leveraging Video Codec Signals for Efficient Streaming VLM Inference

Continuous inference over concurrent video streams imposes substantial compute and memory demands on vision-language model (VLM) serving. Streaming inference uses sliding windows to maintain a bounded context of recent video, but processing each window independently repeats visual encoding and large language model (LLM) prefilling for similar and overlapping content. Existing optimizations provide limited coordination across these stages and often rely on model-specific training, profiling, or model-generated signals. We present CodecSight, a streaming VLM serving system that uses codec metadata as shared runtime guidance across visual encoding and LLM prefilling, without model-specific training or offline profiling. Codec-derived change signals guide patch pruning before visual encoding, reducing both visual computation and the number of downstream visual tokens. Codec-defined frame types guide selective key-value (KV) refresh across windows, while positional correction enables reuse of the remaining cached keys. Across three VLMs and four video workloads, our vLLM-based implementation supports up to $3.3\times$ as many concurrent streams and achieves up to a $5.3\times$ speedup in average time-to-first-token relative to the state-of-the-art baselines. It also reduces executed FLOPs by up to 93%, with a maximum task-quality decrease of 4.64 percentage points.

cs.DC↗

arXiv2026

Rethinking AI Cloud Infrastructure for Agentic Serving Systems with the Aries Experimentation Framework

Autonomous agents challenge conventional LLM serving by coupling repeated inference with persistent context and sandboxed tool execution. We present Aries, a full-stack experimentation framework that separates task semantics from execution configurations, reconstructs cross-component agent trajectories with correlated system telemetry, and exposes stateful tool execution through a consistent interface across heterogeneous sandbox substrates. We use Aries to conduct reproducible experiments on open agent harnesses and benchmarks. We complement these experiments with production traces from a commercial platform, grounding low-level systems research in observed production behavior. Our results show that (1) token-centric metrics miss non-inference bottlenecks, (2) retaining additional context yields diminishing accuracy benefits while reducing serving capacity, and (3) tool sandboxes alternate between long idle periods and short resource bursts, while current snapshot-based state management makes aggressive suspension costly. A complementary security analysis further highlights the need to reduce the sandbox attack surface. We then discuss the vision for agent-native serving systems designed around trajectory-level metrics, adaptive context management, elastic sandbox resource management, and sandboxes with minimized attack surface.

cs.DC↗

arXiv2026

Spotlight: Synergizing Seed Exploration and Spot GPUs for DiT RL Post-Training

Reinforcement learning (RL) post-training of Diffusion Transformers (DiTs) is prohibitively expensive, requiring thousands of high-end GPUs. Existing works explore two directions to reduce cost: seed exploration improves training convergence by selecting high-contrast samples, yet adds compute to the critical path; spot GPUs offer 69--77\% lower cost, yet sit idle during training because DiT rollouts finish nearly simultaneously, which prevents LLM-style pipelining of rollout with training. Spot preemptions further break Sequence Parallelism (SP) groups, fragmenting GPU topology. We present Spotlight, the first system that harvests spot GPUs for DiT RL post-training. Spotlight rests on two key insights we devise: (1)~we show that exploration can tolerate stale model weights because exploration that uses the model weights from the previous iteration preserves the relative ranking of random seeds, allowing exploration to run on idle spot GPUs during training. (2)~SP reconfiguration can reuse on-node state, reducing group recovery from minutes to sub-second launches. Built on these insights, Spotlight introduces three techniques: a bandit-based exploration planner that maximizes reward variance within the training time budget, elastic sequence parallelism that reconfigures SP groups on the fly via persistent schedulers and intra-node weight copying, and a preemption-aware pull-based request scheduler that balances load and commits in-flight state upon preemption. We implement Spotlight on the open-source RL platform ROLL and evaluate it on Qwen-Image post-training. Spotlight reaches the same target validation score $4\times$ faster than baselines, reducing total cost by $1.4$-$6.4\times$ while achieving superior image quality on DeepSeek-OCR and Geneval datasets with resolution $512\times512$ and $1280\times1280$.

cs.DC↗

arXiv2026

FASER: Fine-Grained Phase Management for Speculative Decoding in Dynamic LLM Serving

Speculative decoding (SD) is a widely used approach for accelerating decode-heavy LLM inference workloads. While online inference workloads are highly dynamic, existing SD systems are rigid and take a coarse-grained approach to SD management. They typically set the speculative token length for an entire batch and serialize the execution of the draft and verification phases. Consequently, these systems fall short at adapting to volatile online inference traffic. Under low load, they exhibit prolonged latency because the draft phase blocks the verification phase for the entire batch, leaving GPU computing resources underutilized. Conversely, under high load, they waste computation on rejected tokens during the verification phase, overloading GPU resources. We introduce FASER, a novel system that features fine-grained SD phase management. First, FASER minimizes computational waste by dynamically adjusting the speculative length for each request within a continuous batch and by performing early pruning of rejected tokens inside the verification phase. Second, FASER breaks the verification phase into frontiers, or chunks, to overlap them with the draft phase. This overlap is achieved via fine-grained spatial multiplexing with minimal resource interference. Our FASER prototype in vLLM improves throughput by up to 53% and reduces latency by up to 1.92$\times$ compared to state-of-the-art systems.

cs.DC↗

arXiv2026

Nexus: Transparent I/O Offloading for High-Density Serverless Computing

Serverless computing relies on extreme multi-tenancy to remain economically viable, driving providers to rely on virtual machines (VMs) that ensure strong isolation and seamless ecosystem compatibility with the FaaS programming model. However, current architectures tightly couple application processing logic with I/O processing, forcing every VM to duplicate a heavy communication fabric (cloud SDK, RPC, and TCP/IP). Our analysis reveals this duplication consumes over 25% of a function's memory footprint, and may double the CPU cycles in VMs compared to bare-metal execution. While prior systems attempt to solve this using WebAssembly or library OSes, they naively sacrifice ecosystem compatibility, forcing developers to migrate code and dependencies to new languages. We introduce Nexus, a serverless-native KVM-based hypervisor that transparently decouples compute from I/O. Nexus shifts the execution model by intercepting communication fabric at the API boundary and offloading it to an always-on host shared backend via zero-copy shared memory. This removes the heavyweight communication fabric from the guest VM, while preserving the conventional serverless programming model. By structurally separating these domains, Nexus unlocks asynchronous I/O optimizations: overlapping input payload prefetching with VM restoration from a snapshot and writing output payloads back to storage off the critical path. Compared to the production baseline, Nexus reduces overall node-level CPU and memory consumption by up to 44% and 31%, respectively, thus increasing deployment density by 37%. Also, Nexus reduces warm- and cold-start latency by 39% and 10%, respectively, bringing the response time within 20% of that of a WASM-based, ecosystem-incompatible hypervisor.

cs.DC↗

arXiv2026

PromptTuner: SLO-Aware Elastic System for LLM Prompt Tuning

Prompt tuning has become a prominent strategy for enhancing the performance of Large Language Models (LLMs) on downstream tasks. Many IT enterprises now offer Prompt-Tuning-as-a-Service to fulfill the growing demand for prompt tuning LLMs on downstream tasks. Their primary objective is to satisfy users Service Level Objectives (SLOs) while reducing resource provisioning costs. Nevertheless, our characterization analysis for existing deep learning resource management systems reveals that they are insufficient to optimize these objectives for LLM prompt tuning workloads. In this paper, we introduce PromptTuner, an SLO-aware elastic system to optimize LLM prompt tuning. It contains two innovations. (1) We design a Prompt Bank to identify efficient initial prompts to expedite the convergence of prompt tuning. (2) We develop aWorkload Scheduler to enable fast resource allocation to reduce the SLO violation and resource costs. In our evaluation, PromptTuner reduces SLO violations by 4.0x and 7.9x, and lowers costs by 1.6x and 4.5x, compared to INFless and ElasticFlow respectively.

cs.DC↗

arXiv2026

MemTrust: A Zero-Trust Architecture for Unified AI Memory System

AI memory systems are evolving toward unified context layers that enable efficient cross-agent collaboration and multi-tool workflows, facilitating better accumulation of personal data and learning of user preferences. However, centralization creates a trust crisis where users must entrust cloud providers with sensitive digital memory data. We identify a core tension between personalization demands and data sovereignty: centralized memory systems enable efficient cross-agent collaboration but expose users' sensitive data to cloud provider risks, while private deployments provide security but limit collaboration. To resolve this tension, we aim to achieve local-equivalent security while enabling superior maintenance efficiency and collaborative capabilities. We propose a five-layer architecture abstracting common functional components of AI memory systems: Storage, Extraction, Learning, Retrieval, and Governance. By applying TEE protection to each layer, we establish a trustworthy framework. Based on this, we design MemTrust, a hardware-backed zero-trust architecture that provides cryptographic guarantees across all layers. Our contributions include the five-layer abstraction, "Context from MemTrust" protocol for cross-application sharing, side-channel hardened retrieval with obfuscated access patterns, and comprehensive security analysis. The architecture enables third-party developers to port existing systems with acceptable development costs, achieving system-wide trustworthiness. We believe that AI memory plays a crucial role in enhancing the efficiency and collaboration of agents and AI tools. AI memory will become the foundational infrastructure for AI agents, and MemTrust serves as a universal trusted framework for AI memory systems, with the goal of becoming the infrastructure of memory infrastructure.

cs.CR↗

arXiv2026

Melding the Serverless Control Plane with the Conventional Cluster Manager for Speed and Resource Efficiency

Serverless platforms face a trade-off: conventional cluster managers like Kubernetes offer compatibility for co-locating Function-as-a-Service (FaaS) and Backend-as-a-Service (BaaS) components of serverless applications, at the cost of high cold-start latency, whereas specialized FaaS-only systems like Dirigent achieve low latency by sacrificing compatibility, preventing integrated management and optimization. Our analysis reveals that FaaS traffic is bimodal: predictable, sustainable traffic consumes >98% of cluster resources, whereas sporadic, excessive bursts stress the control plane's scaling latency, not its throughput. With these insights, we design PulseNet, a serverless architecture that uses a dual-track control plane tailored to both traffic types. PulseNet's standard track manages sustainable traffic with long-lived, full-featured Regular Instances under a conventional cluster manager, preserving compatibility for the majority of the workload. To handle excessive traffic, an expedited track bypasses the slow manager to rapidly create short-lived, disposable Emergency Instances, minimizing cold-start latency and resource waste from idle instances. This hybrid approach achieves 35% better performance than Dirigent, a FaaS-only system, on a production workload at the same cost and outperforms other Kubernetes-compatible systems by 1.5-3.5x, reducing the cost by up to 70%.

cs.DC↗

arXiv2025

TokenScale: Timely and Accurate Autoscaling for Disaggregated LLM Serving with Token Velocity

The architectural shift to prefill/decode (PD) disaggregation in LLM serving improves resource utilization but struggles with the bursty nature of modern workloads. Existing autoscaling policies, often retrofitted from monolithic systems like those in AIBrix and DistServe, rely on lagging indicators such as GPU utilization or coarse-grained request counts. This results in slow reactions to load spikes, leading to significant Time-to First-Token (TTFT) and Time-Per-Output-Token (TPOT) SLO violations and costly over-provisioning. We introduce TokenScale, an autoscaling framework that resolves this performance mismatch through two innovations. First, we propose Token Velocity, a novel metric that unifies the prefill, network, and decode stages by quantifying their rate of work. As a leading indicator of system backpressure, it enables proactive scaling. Second, Convertible Decoders allow decoder GPUs to dynamically execute prefill tasks during traffic spikes, creating a rapid-response buffer that absorbs bursts and eliminates the initialization latency of new prefillers. Our evaluation on a GPU cluster with production traces shows TokenScale improves SLO attainment from 50-88% to 80-96% and reduces costs by 4-14% over state-of-the-art systems, including DistServe, BlitzScale, and AIBrix. By uniting a predictive metric with a flexible system design, TokenScale significantly boosts the performance and efficiency of disaggregated LLM serving infrastructure.

cs.DC↗

arXiv2025

Slim-SC: Thought Pruning for Efficient Scaling with Self-Consistency

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.

cs.CL↗

arXiv2025

The High Cost of Keeping Warm: Characterizing Overhead in Serverless Autoscaling Policies

Serverless computing is transforming cloud application development, but the performance-cost trade-offs of control plane designs remain poorly understood due to a lack of open, cross-platform benchmarks and detailed system analyses. In this work, we address these gaps by designing a serverless system that approximates the scaling behaviors of commercial providers, including AWS Lambda and Google Cloud Run. We systematically compare the performance and cost-efficiency of both synchronous and asynchronous autoscaling policies by replaying real-world workloads and varying key autoscaling parameters. We demonstrate that our open-source systems can closely replicate the operational characteristics of commercial platforms, enabling reproducible and transparent experimentation. By evaluating how autoscaling parameters affect latency, memory usage, and CPU overhead, we reveal several key findings. First, we find that serverless systems exhibit significant computational overhead due to instance churn equivalent to 10-40% of the CPU cycles spent on request handling, primarily originating from worker nodes. Second, we observe high memory allocation due to scaling policy: 2-10 times more than actively used. Finally, we demonstrate that reducing these overheads typically results in significant performance degradation in the current systems, underscoring the need for new, cost-efficient autoscaling strategies. Additionally, we employ a hybrid methodology that combines real control plane deployments with large-scale simulation to extend our evaluation closer to a production scale, thereby bridging the gap between small research clusters and real-world environments.

cs.DC↗

arXiv2025

Shattering the Ephemeral Storage Cost Barrier for Data-Intensive Serverless Workflows

Serverless computing is a popular cloud deployment paradigm where developers implement applications as workflows of functions that invoke each other. Cloud providers automatically scale function instances on demand and forward workflow requests to appropriate instances. However, current serverless clouds lack efficient cross-function data transfer, limiting the execution of data-intensive applications. Functions often rely on third-party services like AWS S3, AWS ElastiCache, or multi-tier solutions for intermediate data transfers, which introduces inefficiencies. We demonstrate that such through-storage transfers make data-intensive deployments economically impractical, with storage costs comprising more than 24-99% of the total serverless bill. To address this, we introduce Zipline, a fast, API-preserving data communication method for serverless platforms. Zipline enables direct function-to-function transfers, where the sender function buffers payloads in memory and sends a reference to the receiver. The receiver retrieves the data directly from the sender's memory, guided by the load balancer and autoscaler. Zipline integrates seamlessly with existing autoscaling, maintains invocation semantics, and eliminates the costs and overheads of intermediate services. We prototype Zipline in vHive/Knative on AWS EC2 nodes, demonstrating significant improvements. Zipline reduces costs and enhances latency and bandwidth compared to AWS S3 (the lowest-cost solution) and ElastiCache (the highest-performance solution). On real-world applications, Zipline lowers costs by 2-5x and reduces execution times by 1.3-3.4x versus S3. Compared to ElastiCache, Zipline achieves 17-772x cost reductions while improving performance by 2-5%.

cs.DC↗

arXiv2024

ServerlessLLM: Low-Latency Serverless Inference for Large Language Models

This paper presents ServerlessLLM, a distributed system designed to support low-latency serverless inference for Large Language Models (LLMs). By harnessing the substantial near-GPU storage and memory capacities of inference servers, ServerlessLLM achieves effective local checkpoint storage, minimizing the need for remote checkpoint downloads and ensuring efficient checkpoint loading. The design of ServerlessLLM features three core contributions: (i) \emph{fast multi-tier checkpoint loading}, featuring a new loading-optimized checkpoint format and a multi-tier loading system, fully utilizing the bandwidth of complex storage hierarchies on GPU servers; (ii) \emph{efficient live migration of LLM inference}, which enables newly initiated inferences to capitalize on local checkpoint storage while ensuring minimal user interruption; and (iii) \emph{startup-time-optimized model scheduling}, which assesses the locality statuses of checkpoints on each server and schedules the model onto servers that minimize the time to start the inference. Comprehensive evaluations, including microbenchmarks and real-world scenarios, demonstrate that ServerlessLLM dramatically outperforms state-of-the-art serverless systems, reducing latency by 10 - 200X across various LLM inference workloads.

cs.LG↗

arXiv2021

Benchmarking, Analysis, and Optimization of Serverless Function Snapshots

Serverless computing has seen rapid adoption due to its high scalability and flexible, pay-as-you-go billing model. In serverless, developers structure their services as a collection of functions, sporadically invoked by various events like clicks. High inter-arrival time variability of function invocations motivates the providers to start new function instances upon each invocation, leading to significant cold-start delays that degrade user experience. To reduce cold-start latency, the industry has turned to snapshotting, whereby an image of a fully-booted function is stored on disk, enabling a faster invocation compared to booting a function from scratch. This work introduces vHive, an open-source framework for serverless experimentation with the goal of enabling researchers to study and innovate across the entire serverless stack. Using vHive, we characterize a state-of-the-art snapshot-based serverless infrastructure, based on industry-leading Containerd orchestration framework and Firecracker hypervisor technologies. We find that the execution time of a function started from a snapshot is 95% higher, on average, than when the same function is memory-resident. We show that the high latency is attributable to frequent page faults as the function's state is brought from disk into guest memory one page at a time. Our analysis further reveals that functions access the same stable working set of pages across different invocations of the same function. By leveraging this insight, we build REAP, a light-weight software mechanism for serverless hosts that records functions' stable working set of guest memory pages and proactively prefetches it from disk into memory. Compared to baseline snapshotting, REAP slashes the cold-start delays by 3.7x, on average.

cs.DC↗

arXiv2020

Bankrupt Covert Channel: Turning Network Predictability into Vulnerability

Recent years have seen a surge in the number of data leaks despite aggressive information-containment measures deployed by cloud providers. When attackers acquire sensitive data in a secure cloud environment, covert communication channels are a key tool to exfiltrate the data to the outside world. While the bulk of prior work focused on covert channels within a single CPU, they require the spy (transmitter) and the receiver to share the CPU, which might be difficult to achieve in a cloud environment with hundreds or thousands of machines. This work presents Bankrupt, a high-rate highly clandestine channel that enables covert communication between the spy and the receiver running on different nodes in an RDMA network. In Bankrupt, the spy communicates with the receiver by issuing RDMA network packets to a private memory region allocated to it on a different machine (an intermediary). The receiver similarly allocates a separate memory region on the same intermediary, also accessed via RDMA. By steering RDMA packets to a specific set of remote memory addresses, the spy causes deep queuing at one memory bank, which is the finest addressable internal unit of main memory. This exposes a timing channel that the receiver can listen on by issuing probe packets to addresses mapped to the same bank but in its own private memory region. Bankrupt channel delivers 74Kb/s throughput in CloudLab's public cloud while remaining undetectable to the existing monitoring capabilities, such as CPU and NIC performance counters.

cs.CR↗

arXiv2019

Design Guidelines for High-Performance SCM Hierarchies

With emerging storage-class memory (SCM) nearing commercialization, there is evidence that it will deliver the much-anticipated high density and access latencies within only a few factors of DRAM. Nevertheless, the latency-sensitive nature of memory-resident services makes seamless integration of SCM in servers questionable. In this paper, we ask the question of how best to introduce SCM for such servers to improve overall performance/cost over existing DRAM-only architectures. We first show that even with the most optimistic latency projections for SCM, the higher memory access latency results in prohibitive performance degradation. However, we find that deployment of a modestly sized high-bandwidth 3D stacked DRAM cache makes the performance of an SCM-mostly memory system competitive. The high degree of spatial locality that memory-resident services exhibit not only simplifies the DRAM cache's design as page-based, but also enables the amortization of increased SCM access latencies and the mitigation of SCM's read/write latency disparity. We identify the set of memory hierarchy design parameters that plays a key role in the performance and cost of a memory system combining an SCM technology and a 3D stacked DRAM cache. We then introduce a methodology to drive provisioning for each of these design parameters under a target performance/cost goal. Finally, we use our methodology to derive concrete results for specific SCM technologies. With PCM as a case study, we show that a two bits/cell technology hits the performance/cost sweet spot, reducing the memory subsystem cost by 40% while keeping performance within 3% of the best performing DRAM-only system, whereas single-level and triple-level cell organizations are impractical for use as memory replacements.

cs.AR↗