arXiv subjects

Francisco Romero

Publications and source records attributed to Francisco Romero.

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

Halo: Domain-Aware Query Optimization for Long-Context Question Answering

Long-context question answering (QA) over lengthy documents is critical for applications such as financial analysis, legal review, and scientific research. Current approaches, such as processing entire documents via a single LLM call or retrieving relevant chunks via RAG have two drawbacks: First, as context size increases, response quality can degrade, impacting accuracy. Second, iteratively processing hundreds of input documents can incur prohibitively high costs in API calls. To improve response quality and reduce the number of iterations needed to get the desired response, users tend to add domain knowledge to their prompts. However, existing systems fail to systematically capture and use this knowledge to guide query processing. Domain knowledge is treated as prompt tokens alongside the document: the LLM may or may not follow it, there is no reduction in computational cost, and when outputs are incorrect, users must manually iterate. We present Halo, a long-context QA framework that automatically extracts domain knowledge from user prompts and applies it as executable operators across a multi-stage query execution pipeline. Halo identifies three common forms of domain knowledge - where in the document to look, what content to ignore, and how to verify the answer - and applies each at the pipeline stage where it is most effective: pruning the document before chunk selection, filtering irrelevant chunks before inference, and ranking candidate responses after generation. To handle imprecise or invalid domain knowledge, Halo includes a fallback mechanism that detects low-quality operators at runtime and selectively disables them. Our evaluation across finance, literature, and scientific datasets shows that Halo achieves up to 13% higher accuracy and 4.8x lower cost compared to baselines, and enables a lightweight open-source model to approach frontier LLM accuracy at 78x lower cost.

cs.DB↗

arXiv2025

Flash-Fusion: Enabling Expressive, Low-Latency Queries on IoT Sensor Streams with LLMs

Smart cities and pervasive IoT deployments have generated interest in IoT data analysis across transportation and urban planning. At the same time, Large Language Models offer a new interface for exploring IoT data - particularly through natural language. Users today face two key challenges when working with IoT data using LLMs: (1) data collection infrastructure is expensive, producing terabytes of low-level sensor readings that are too granular for direct use, and (2) data analysis is slow, requiring iterative effort and technical expertise. Directly feeding all IoT telemetry to LLMs is impractical due to finite context windows, prohibitive token costs at scale, and non-interactive latencies. What is missing is a system that first parses a user's query to identify the analytical task, then selects the relevant data slices, and finally chooses the right representation before invoking an LLM. We present Flash-Fusion, an end-to-end edge-cloud system that reduces the IoT data collection and analysis burden on users. Two principles guide its design: (1) edge-based statistical summarization (achieving 73.5% data reduction) to address data volume, and (2) cloud-based query planning that clusters behavioral data and assembles context-rich prompts to address data interpretation. We deploy Flash-Fusion on a university bus fleet and evaluate it against a baseline that feeds raw data to a state-of-the-art LLM. Flash-Fusion achieves a 95% latency reduction and 98% decrease in token usage and cost while maintaining high-quality responses. It enables personas across disciplines - safety officers, urban planners, fleet managers, and data scientists - to efficiently iterate over IoT data without the burden of manual query authoring or preprocessing.

cs.DC↗

arXiv2023

Zelda: Video Analytics using Vision-Language Models

Advances in ML have motivated the design of video analytics systems that allow for structured queries over video datasets. However, existing systems limit query expressivity, require users to specify an ML model per predicate, rely on complex optimizations that trade off accuracy for performance, and return large amounts of redundant and low-quality results. This paper focuses on the recently developed Vision-Language Models (VLMs) that allow users to query images using natural language like "cars during daytime at traffic intersections." Through an in-depth analysis, we show VLMs address three limitations of current video analytics systems: general expressivity, a single general purpose model to query many predicates, and are both simple and fast. However, VLMs still return large numbers of redundant and low-quality results that can overwhelm and burden users. In addition, VLMs often require manual prompt engineering to improve result relevance. We present Zelda: a video analytics system that uses VLMs to return both relevant and semantically diverse results for top-K queries on large video datasets. Zelda prompts the VLM with the user's query in natural language. Zelda then automatically adds discriminator and synonym terms to boost accuracy, and terms to identify low-quality frames. To improve result diversity, Zelda uses semantic-rich VLM embeddings in an algorithm that prunes similar frames while considering their relevance to the query and the number of top-K results requested. We evaluate Zelda across five datasets and 19 queries and quantitatively show it achieves higher mean average precision (up to 1.15x) and improves average pairwise similarity (up to 1.16x) compared to using VLMs out-of-the-box. We also compare Zelda to a state-of-the-art video analytics engine and show that Zelda retrieves results 7.5x (up to 10.4x) faster for the same accuracy and frame diversity.

cs.DB↗

arXiv2022

A generalized conditional gradient method for dynamic inverse problems with optimal transport regularization

We develop a dynamic generalized conditional gradient method (DGCG) for dynamic inverse problems with optimal transport regularization. We consider the framework introduced in (Bredies and Fanzon, ESAIM: M2AN, 54:2351-2382, 2020), where the objective functional is comprised of a fidelity term, penalizing the pointwise in time discrepancy between the observation and the unknown in time-varying Hilbert spaces, and a regularizer keeping track of the dynamics, given by the Benamou-Brenier energy constrained via the homogeneous continuity equation. Employing the characterization of the extremal points of the Benamou-Brenier energy (Bredies et al., arXiv:1907.11589, 2019) we define the atoms of the problem as measures concentrated on absolutely continuous curves in the domain. We propose a dynamic generalization of a conditional gradient method that consists in iteratively adding suitably chosen atoms to the current sparse iterate, and subsequently optimize the coefficients in the resulting linear combination. We prove that the method converges with a sublinear rate to a minimizer of the objective functional. Additionally, we propose heuristic strategies and acceleration steps that allow to implement the algorithm efficiently. Finally, we provide numerical examples that demonstrate the effectiveness of our algorithm and model at reconstructing heavily undersampled dynamic data, together with the presence of noise.

math.NA↗

arXiv2021

Llama: A Heterogeneous & Serverless Framework for Auto-Tuning Video Analytics Pipelines

The proliferation of camera-enabled devices and large video repositories has led to a diverse set of video analytics applications. These applications rely on video pipelines, represented as DAGs of operations, to transform videos, process extracted metadata, and answer questions like, "Is this intersection congested?" The latency and resource efficiency of pipelines can be optimized using configurable knobs for each operation (e.g., sampling rate, batch size, or type of hardware used). However, determining efficient configurations is challenging because (a) the configuration search space is exponentially large, and (b) the optimal configuration depends on users' desired latency and cost targets, (c) input video contents may exercise different paths in the DAG and produce a variable amount intermediate results. Existing video analytics and processing systems leave it to the users to manually configure operations and select hardware resources. We present Llama: a heterogeneous and serverless framework for auto-tuning video pipelines. Given an end-to-end latency target, Llama optimizes for cost efficiency by (a) calculating a latency target for each operation invocation, and (b) dynamically running a cost-based optimizer to assign configurations across heterogeneous hardware that best meet the calculated per-invocation latency target. This makes the problem of auto-tuning large video pipelines tractable and allows us to handle input-dependent behavior, conditional branches in the DAG, and execution variability. We describe the algorithms in Llama and evaluate it on a cloud platform using serverless CPU and GPU resources. We show that compared to state-of-the-art cluster and serverless video analytics and processing systems, Llama achieves 7.8x lower latency and 16x cost reduction on average.

cs.DC↗

arXiv2021

Faa$T: A Transparent Auto-Scaling Cache for Serverless Applications

Function-as-a-Service (FaaS) has become an increasingly popular way for users to deploy their applications without the burden of managing the underlying infrastructure. However, existing FaaS platforms rely on remote storage to maintain state, limiting the set of applications that can be run efficiently. Recent caching work for FaaS platforms has tried to address this problem, but has fallen short: it disregards the widely different characteristics of FaaS applications, does not scale the cache based on data access patterns, or requires changes to applications. To address these limitations, we present Faa\$T, a transparent auto-scaling distributed cache for serverless applications. Each application gets its own Faa\$T cache. After a function executes and the application becomes inactive, the cache is unloaded from memory with the application. Upon reloading for the next invocation, Faa\$T pre-warms the cache with objects likely to be accessed. In addition to traditional compute-based scaling, Faa\$T scales based on working set and object sizes to manage cache space and I/O bandwidth. We motivate our design with a comprehensive study of data access patterns in a large-scale commercial FaaS provider. We implement Faa\$T for the provider's production FaaS platform. Our experiments show that Faa\$T can improve performance by up to 92% (57% on average) for challenging applications, and reduce cost for most users compared to state-of-the-art caching systems, i.e. the cost of having to stand up additional serverful resources.

cs.DC↗

arXiv2020

INFaaS: A Model-less and Managed Inference Serving System

Despite existing work in machine learning inference serving, ease-of-use and cost efficiency remain challenges at large scales. Developers must manually search through thousands of model-variants -- versions of already-trained models that differ in hardware, resource footprints, latencies, costs, and accuracies -- to meet the diverse application requirements. Since requirements, query load, and applications themselves evolve over time, these decisions need to be made dynamically for each inference query to avoid excessive costs through naive autoscaling. To avoid navigating through the large and complex trade-off space of model-variants, developers often fix a variant across queries, and replicate it when load increases. However, given the diversity across variants and hardware platforms in the cloud, a lack of understanding of the trade-off space can incur significant costs to developers. This paper introduces INFaaS, a managed and model-less system for distributed inference serving, where developers simply specify the performance and accuracy requirements for their applications without needing to specify a specific model-variant for each query. INFaaS generates model-variants, and efficiently navigates the large trade-off space of model-variants on behalf of developers to meet application-specific objectives: (a) for each query, it selects a model, hardware architecture, and model optimizations, (b) it combines VM-level horizontal autoscaling with model-level autoscaling, where multiple, different model-variants are used to serve queries within each machine. By leveraging diverse variants and sharing hardware resources across models, INFaaS achieves 1.3x higher throughput, violates latency objectives 1.6x less often, and saves up to 21.6x in cost (8.5x on average) compared to state-of-the-art inference serving systems on AWS EC2.

cs.DC↗

arXiv2020

On the extremal points of the ball of the Benamou-Brenier energy

In this paper we characterize the extremal points of the unit ball of the Benamou--Brenier energy and of a coercive generalization of it, both subjected to the homogeneous continuity equation constraint. We prove that extremal points consist of pairs of measures concentrated on absolutely continuous curves which are characteristics of the continuity equation. Then, we apply this result to provide a representation formula for sparse solutions of dynamic inverse problems with finite dimensional data and optimal-transport based regularization.

math.OC↗

arXiv2020

The TRaCaR Ratio: Selecting the Right Storage Technology for Active Dataset-Serving Databases

Main memory database systems aim to provide users with low latency and high throughput access to data. Most data resides in secondary storage, which is limited by the access speed of the technology. For hot content, data resides in DRAM, which has become increasingly expensive as datasets grow in size and access demand. With the emergence of low-latency storage solutions such as Flash and Intel's 3D XPoint (3DXP), there is an opportunity for these systems to give users high Quality-of-Service while reducing the cost for providers. To achieve high performance, providers must provision the server hosts for these datasets with the proper amount of DRAM and secondary storage, as well as selecting a storage technology. The growth of capacity and transaction load overtime makes it expensive to flip back-and-forth between different storage technologies and memory-storage combinations. Servers set up for one storage technology must now be reconfigured, repartitioned, and potentially replaced altogether. As more low-latency storage solutions become available, how does one decide on the right memory-storage combination, as well as selecting a storage technology, given a predicted trend in dataset growth and offered load? In this paper, we describe and make the case for using the TRaCaR ratio - the transaction rate divided by the storage capacity needed for a workload - for allowing providers to choose the most cost-effective memory-storage combination and storage technology given their predicted dataset trend and load requirement. We explore how the TRaCaR ratio can be used with 3DXP and Flash with a highly-zipfian b-tree database, and discuss potential research directions that can leverage the ratio.

cs.DC↗

arXiv2018

Dynamic Spike Super-resolution and Applications to Ultrafast Ultrasound Imaging

We consider the dynamical super-resolution problem consisting in the recovery of positions and velocities of moving particles from low-frequency static measurements taken over multiple time steps. The standard approach to this issue is a two-step process: first, at each time step some static reconstruction method is applied to locate the positions of the particles with super-resolution and, second, some tracking technique is applied to obtain the velocities. In this paper we propose a fully dynamical method based on a phase-space lifting of the positions and the velocities of the particles, which are simultaneously reconstructed with super-resolution. We provide a rigorous mathematical analysis of the recovery problem, both for the noiseless case and in presence of noise (in the discrete setting). Several numerical simulations illustrate and validate our method, which shows some advantage over existing techniques. We then discuss the application of this approach to the dynamical super-resolution problem in ultrafast ultrasound imaging: blood vessels' locations and blood flow velocities are recovered with super-resolution.

math.NA↗

arXiv2018

Mage: Online Interference-Aware Scheduling in Multi-Scale Heterogeneous Systems

Heterogeneity has grown in popularity both at the core and server level as a way to improve both performance and energy efficiency. However, despite these benefits, scheduling applications in heterogeneous machines remains challenging. Additionally, when these heterogeneous resources accommodate multiple applications to increase utilization, resources are prone to contention, destructive interference, and unpredictable performance. Existing solutions examine heterogeneity either across or within a server, leading to missed performance and efficiency opportunities. We present Mage, a practical interference-aware runtime that optimizes performance and efficiency in systems with intra- and inter-server heterogeneity. Mage leverages fast and online data mining to quickly explore the space of application placements, and determine the one that minimizes destructive interference between co-resident applications. Mage continuously monitors the performance of active applications, and, upon detecting QoS violations, it determines whether alternative placements would prove more beneficial, taking into account any overheads from migration. Across 350 application mixes on a heterogeneous CMP, Mage improves performance by 38% and up to 2x compared to a greedy scheduler. Across 160 mixes on a heterogeneous cluster, Mage improves performance by 30% on average and up to 52% over the greedy scheduler, and by 11% over the combination of Paragon [15] for inter- and intra-server heterogeneity.

cs.DC↗

arXiv2017

Heat generation with plasmonic nanoparticles

In this paper we use layer potentials and asymptotic analysis techniques to analyze the heat generation due to nanoparticles when illuminated at their plasmonic resonance. We consider arbitrary-shaped particles and both single and multiple particles. For close-to-touching nanoparticles, we show that the temperature field deviates significantly from the one generated by a single nanoparticle. The results of this paper open a door for solving the challenging problems of detecting plasmonic nanoparticles in biological media and monitoring temperature elevation in tissue generated by nanoparticle heating.

math.AP↗

arXiv2016

Mathematical Analysis of Ultrafast Ultrasound Imaging

This paper provides a mathematical analysis of ultrafast ultrasound imaging. This newly emerging modality for biomedical imaging uses plane waves instead of focused waves in order to achieve very high frame rates. We derive the point spread function of the system in the Born approximation for wave propagation and study its properties. We consider dynamic data for blood flow imaging, and introduce a suitable random model for blood cells. We show that a singular value decomposition method can successfully remove the clutter signal by using the different spatial coherence of tissue and blood signals, thereby providing high-resolution images of blood vessels, even in cases when the clutter and blood speeds are comparable in magnitude. Several numerical simulations are presented to illustrate and validate the approach.

math.NA↗

arXiv2016

A signal separation technique for sub-cellular imaging using dynamic optical coherence tomography

This paper aims at imaging the dynamics of metabolic activity of cells. Using dynamic optical coherence tomography, we introduce a new multi-particle dynamical model to simulate the movements of the collagen and the cell metabolic activity and develop an efficient signal separation technique for sub-cellular imaging. We perform a singular-value decomposition of the dynamic optical images to isolate the intensity of the metabolic activity. We prove that the largest eigenvalue of the associated Casorati matrix corresponds to the collagen. We present several numerical simulations to illustrate and validate our approach.

math.NA↗

arXiv2015

Simultaneous source and attenuation reconstruction in SPECT using ballistic and single scattering data

In medical SPECT imaging, we seek to simultaneously obtain the internal radioactive sources and the attenuation map using not only ballistic measurements but also first order scattering measurements. The problem is modeled using the radiative transfer equation by means of an explicit nonlinear operator that gives the ballistic and scattering measurements as a function of the radioactive source and attenuation distributions. First, by differentiating this nonlinear operator we obtain a linearized inverse problem. Then, under regularity hypothesis for the source distribution and attenuation map and considering small attenuations, we rigorously prove that the linear operator is invertible and we compute its inverse explicitly. This allows to prove local uniqueness for the nonlinear inverse problem. Finally, using the previous inversion result for the linear operator, we propose a new type of iterative algorithm for simultaneous source and attenuation recovery for SPECT based on Neumann series and a Newton-Raphson algorithm.

math.AP↗