arXiv subjects

Nishil Talati

Publications and source records attributed to Nishil Talati.

At least 19 recordsLinked to original sources

arXiv2026

Energy Calculus: A Compositional Algebra of Energy in Computational Systems

Energy is a binding constraint for AI scaling, yet it lacks the formal treatment that computation, communication, and learning have long enjoyed. Recent systems demonstrate large energy savings, but each targets a specific granularity and structure; one cannot combine frequency scaling from one system with critical-path analysis from another and reason about their joint effect on total energy. Energy remains a monolithic scalar that is measured after the fact and optimized with point solutions that do not generalize. We propose energy calculus, a compositional algebra that treats energy as a first-class primitive. It builds on energy elements, units of computation whose energy we can reliably measure, each carrying an energy signature that comprises its time, its static and dynamic energy, the hardware operating point and execution context under which we measured it, and the associated measurement uncertainty. Three operators (sequential, same-device parallel, and cross-device parallel) compose signatures along the same structure as the computation itself, covering arbitrary DAG-structured executions. The algebra rests on seven axioms that capture how hardware consumes energy, and it exhibits two properties distinctive to energy among computing resources: sequential composition commutes only when elements are mutually context-insensitive, and sequential composition does not distribute over parallel composition. We also present a Reduction Theorem that recovers simple context-independent algebra whenever interactions fall below measurement uncertainty, so practitioners pay for context dependence only where the physics demands it. Uncertainty propagates through every composition, so each prediction carries an error bound. Finally, we show that the same operators extend from energy totals to time--energy Pareto frontiers, so reasoning about tradeoffs composes with the same algebra.

cs.DC↗

arXiv2026

Agentic AI Workload Characteristics

Agentic AI shifts LLM serving from isolated prompt-generation requests to stateful, multi-turn executions that repeatedly invoke the model, call tools, and grow context over time. This paper characterizes ReAct-style agents from both the LLM-serving and tool-execution perspectives using an end-to-end tracing infrastructure across reasoning and non-reasoning Gemma and Qwen configurations on five agentic benchmarks. Our study shows that agentic workloads are not simply long-prompt workloads: with effective context caching, most input tokens are reused across turns, making execution decode-dominated while increasing dependence on long-lived KV-cache state. We also find that tool use has a clear temporal structure, with agents shifting from read/explore behavior early in execution to execute/write behavior later. These results show that efficient agentic serving must jointly manage repeated model re-entry, persistent context state, and workload-dependent tool behavior.

cs.DC↗

arXiv2026

KAIROS: Stateful, Context-Aware Power-Efficient Agentic Inference Serving

Power has become a central bottleneck for AI inference. This problem is becoming more urgent as agentic AI emerges as a major workload class, yet prior power-management techniques focus almost entirely on single-turn LLM serving. Our analysis shows that agentic serving behaves fundamentally differently: each request carries long-lived context that evolves across tool-interleaved turns, and lowering GPU frequency can push the system into a thrashing regime where memory pressure sharply worsens both performance and power efficiency. These observations show that power optimization for agentic serving requires rethinking. We present KAIROS, a context-aware power optimization system for agentic AI serving. KAIROS uses agent context as a first-class control signal to jointly manage GPU frequency, per-instance concurrency, and multi-instance request placement. This enables KAIROS to save power when memory headroom exists while avoiding thrashing and preserving performance targets. At a high level, KAIROS tracks requests at agent granularity, adapts local control to context growth and agent progress, and routes agents across instances to jointly improve power efficiency and memory stability. Evaluated across diverse software and data engineering agentic tasks, KAIROS achieves an average of 27% (up to 39.8%) power reduction while meeting the performance targets.

cs.DC↗

arXiv2026

BlazingAML: High-Throughput Anti-Money Laundering (AML) via Multi-Stage Graph Mining

Money laundering detection faces challenges due to excessive false positives and inadequate adaptation to sophisticated multi-stage schemes that exploit modern financial networks. Graph analytics and AI are promising tools, but they struggle with the fuzziness of laundering patterns, which exhibit structural and temporal variations. Conventional data mining techniques require the detailed enumeration of pattern variants, which not only complicates the analyst's task to specify them, but also leads to large run-time overheads and difficulty training accurate AI models. The paper presents BlazingAML, a scalable AML system design that introduces: 1. A novel multi-stage framework for expressing fuzzy money laundering patterns 2. A domain-specific compiler that transforms high-level pattern descriptions into high-performance code for CPU and GPU back-ends The multi-stage abstraction decomposes complex laundering schemes into logical stages connected by graph operations, enabling diverse patterns to be expressed using unified primitives while capturing structural and temporal fuzziness. The compiler applies sophisticated optimizations, eliminating manual parallel programming requirements for financial analysts. Evaluation on IBM AML datasets shows BlazingAML achieves the same F1 score as state-of-the-art approaches while delivering 210x and 333x higher speedup on CPU and GPU respectively, with superior scalability.

cs.DC↗

arXiv2025

Sphinx: Efficiently Serving Novel View Synthesis using Regression-Guided Selective Refinement

Novel View Synthesis (NVS) is the task of generating new images of a scene from viewpoints that were not part of the original input. Diffusion-based NVS can generate high-quality, temporally consistent images, however, remains computationally prohibitive. Conversely, regression-based NVS offers suboptimal generation quality despite requiring significantly lower compute; leaving the design objective of a high-quality, inference-efficient NVS framework an open challenge. To close this critical gap, we present Sphinx, a training-free hybrid inference framework that achieves diffusion-level fidelity at a significantly lower compute. Sphinx proposes to use regression-based fast initialization to guide and reduce the denoising workload for the diffusion model. Additionally, it integrates selective refinement with adaptive noise scheduling, allowing more compute to uncertain regions and frames. This enables Sphinx to provide flexible navigation of the performance-quality trade-off, allowing adaptation to latency and fidelity requirements for dynamically changing inference scenarios. Our evaluation shows that Sphinx achieves an average 1.8x speedup over diffusion model inference with negligible perceptual degradation of less than 5%, establishing a new Pareto frontier between quality and latency in NVS serving.

cs.CV↗

arXiv2025

ZKProphet: Understanding Performance of Zero-Knowledge Proofs on GPUs

Zero-Knowledge Proofs (ZKP) are protocols which construct cryptographic proofs to demonstrate knowledge of a secret input in a computation without revealing any information about the secret. ZKPs enable novel applications in private and verifiable computing such as anonymized cryptocurrencies and blockchain scaling and have seen adoption in several real-world systems. Prior work has accelerated ZKPs on GPUs by leveraging the inherent parallelism in core computation kernels like Multi-Scalar Multiplication (MSM). However, we find that a systematic characterization of execution bottlenecks in ZKPs, as well as their scalability on modern GPU architectures, is missing in the literature. This paper presents ZKProphet, a comprehensive performance study of Zero-Knowledge Proofs on GPUs. Following massive speedups of MSM, we find that ZKPs are bottlenecked by kernels like Number-Theoretic Transform (NTT), as they account for up to 90% of the proof generation latency on GPUs when paired with optimized MSM implementations. Available NTT implementations under-utilize GPU compute resources and often do not employ architectural features like asynchronous compute and memory operations. We observe that the arithmetic operations underlying ZKPs execute exclusively on the GPU's 32-bit integer pipeline and exhibit limited instruction-level parallelism due to data dependencies. Their performance is thus limited by the available integer compute units. While one way to scale the performance of ZKPs is adding more compute units, we discuss how runtime parameter tuning for optimizations like precomputed inputs and alternative data representations can extract additional speedup. With this work, we provide the ZKP community a roadmap to scale performance on GPUs and construct definitive GPU-accelerated ZKPs for their application requirements and available hardware resources.

cs.DC↗

arXiv2025

TIMEST: Temporal Information Motif Estimator Using Sampling Trees

The mining of pattern subgraphs, known as motifs, is a core task in the field of graph mining. Edges in real-world networks often have timestamps, so there is a need for temporal motif mining. A temporal motif is a richer structure that imposes timing constraints on the edges of the motif. Temporal motifs have been used to analyze social networks, financial transactions, and biological networks. Motif counting in temporal graphs is particularly challenging. A graph with millions of edges can have trillions of temporal motifs, since the same edge can occur with multiple timestamps. There is a combinatorial explosion of possibilities, and state-of-the-art algorithms cannot manage motifs with more than four vertices. In this work, we present TIMEST: a general, fast, and accurate estimation algorithm to count temporal motifs of arbitrary sizes in temporal networks. Our approach introduces a temporal spanning tree sampler that leverages weighted sampling to generate substructures of target temporal motifs. This method carefully takes a subset of temporal constraints of the motif that can be jointly and efficiently sampled. TIMEST uses randomized estimation techniques to obtain accurate estimates of motif counts. We give theoretical guarantees on the running time and approximation guarantees of TIMEST. We perform an extensive experimental evaluation and show that TIMEST is both faster and more accurate than previous algorithms. Our CPU implementation exhibits an average speedup of 28x over state-of-the-art GPU implementation of the exact algorithm, and 6x speedup over SOTA approximate algorithms while consistently showcasing less than 5% error in most cases. For example, TIMEST can count the number of instances of a financial fraud temporal motif in four minutes with 0.6% error, while exact methods take more than two days.

cs.DB↗

arXiv2025

Triadic First-Order Logic Queries in Temporal Networks

Motif counting is a fundamental problem in network analysis, and there is a rich literature of theoretical and applied algorithms for this problem. Given a large input network $G$, a motif $H$ is a small "pattern" graph indicative of special local structure. Motif/pattern mining involves finding all matches of this pattern in the input $G$. The simplest, yet challenging, case of motif counting is when $H$ has three vertices, often called a "triadic" query. Recent work has focused on "temporal graph mining", where the network $G$ has edges with timestamps (and directions) and $H$ has time constraints. Inspired by concepts in logic and database theory, we introduce the study of "thresholded First Order Logic (FOL) Motif Analysis" for massive temporal networks. A typical triadic motif query asks for the existence of three vertices that form a desired temporal pattern. An "FOL" motif query is obtained by having both existential and thresholded universal quantifiers. This allows for query semantics that can mine richer information from networks. A typical triadic query would be "find all triples of vertices $u,v,w$ such that they form a triangle within one hour". A thresholded FOL query can express "find all pairs $u,v$ such that for half of $w$ where $(u,w)$ formed an edge, $(v,w)$ also formed an edge within an hour". We design the first algorithm, FOLTY, for mining thresholded FOL triadic queries. The theoretical running time of FOLTY matches the best known running time for temporal triangle counting in sparse graphs. We give an efficient implementation of FOLTY using specialized temporal data structures. FOLTY has excellent empirical behavior, and can answer triadic FOL queries on graphs with nearly 70M edges is less than hour on commodity hardware. Our work has the potential to start a new research direction in the classic well-studied problem of motif analysis.

cs.DB↗

arXiv2025

Mayura: Exploiting Similarities in Motifs for Temporal Co-Mining

Temporal graphs serve as a critical foundation for modeling evolving interactions in domains ranging from financial networks to social media. Mining temporal motifs is essential for applications such as fraud detection, cybersecurity, and dynamic network analysis. However, conventional motif mining approaches treat each query independently, incurring significant redundant computations when similar substructures exist across multiple motifs. In this paper, we propose Mayura, a novel framework that unifies the mining of multiple temporal motifs by exploiting their inherent structural and temporal commonalities. Central to our approach is the Motif-Group Tree (MG-Tree), a hierarchical data structure that organizes related motifs and enables the reuse of common search paths, thereby reducing redundant computation. We propose a co-mining algorithm that leverages the MG-Tree and develop a flexible runtime capable of exploiting both CPU and GPU architectures for scalable performance. Empirical evaluations on diverse real-world datasets demonstrate that Mayura achieves substantial improvements over the state-of-the-art techniques that mine each motif individually, with an average speed-up of 2.4x on the CPU and 1.7x on the GPU, while maintaining the exactness required for high-stakes applications.

cs.DB↗

arXiv2025

DX100: A Programmable Data Access Accelerator for Indirection

Indirect memory accesses frequently appear in applications where memory bandwidth is a critical bottleneck. Prior indirect memory access proposals, such as indirect prefetchers, runahead execution, fetchers, and decoupled access/execute architectures, primarily focus on improving memory access latency by loading data ahead of computation but still rely on the DRAM controllers to reorder memory requests and enhance memory bandwidth utilization. DRAM controllers have limited visibility to future memory accesses due to the small capacity of request buffers and the restricted memory-level parallelism of conventional core and memory systems. We introduce DX100, a programmable data access accelerator for indirect memory accesses. DX100 is shared across cores to offload bulk indirect memory accesses and associated address calculation operations. DX100 reorders, interleaves, and coalesces memory requests to improve DRAM row-buffer hit rate and memory bandwidth utilization. DX100 provides a general-purpose ISA to support diverse access types, loop patterns, conditional accesses, and address calculations. To support this accelerator without significant programming efforts, we discuss a set of MLIR compiler passes that automatically transform legacy code to utilize DX100. Experimental evaluations on 12 benchmarks spanning scientific computing, database, and graph applications show that DX100 achieves performance improvements of 2.6x over a multicore baseline and 2.0x over the state-of-the-art indirect prefetcher.

cs.AR↗

arXiv2025

NMP-PaK: Near-Memory Processing Acceleration of Scalable De Novo Genome Assembly

De novo assembly enables investigations of unknown genomes, paving the way for personalized medicine and disease management. However, it faces immense computational challenges arising from the excessive data volumes and algorithmic complexity. While state-of-the-art de novo assemblers utilize distributed systems for extreme-scale genome assembly, they demand substantial computational and memory resources. They also fail to address the inherent challenges of de novo assembly, including a large memory footprint, memory-bound behavior, and irregular data patterns stemming from complex, interdependent data structures. Given these challenges, de novo assembly merits a custom hardware solution, though existing approaches have not fully addressed the limitations. We propose NMP-PaK, a hardware-software co-design that accelerates scalable de novo genome assembly through near-memory processing (NMP). Our channel-level NMP architecture addresses memory bottlenecks while providing sufficient scratchpad space for processing elements. Customized processing elements maximize parallelism while efficiently handling large data structures that are both dynamic and interdependent. Software optimizations include customized batch processing to reduce the memory footprint and hybrid CPU-NMP processing to address hardware underutilization caused by irregular data patterns. NMP-PaK conducts the same genome assembly while incurring a 14X smaller memory footprint compared to the state-of-the-art de novo assembly. Moreover, NMP-PaK delivers a 16X performance improvement over the CPU baseline, with a 2.4X reduction in memory operations. Consequently, NMP-PaK achieves 8.3X greater throughput than state-of-the-art de novo assembly under the same resource constraints, showcasing its superior computational efficiency.

cs.AR↗

arXiv2025

MoE-Lens: Towards the Hardware Limit of High-Throughput MoE LLM Serving Under Resource Constraints

Mixture of Experts (MoE) LLMs, characterized by their sparse activation patterns, offer a promising approach to scaling language models while avoiding proportionally increasing the inference cost. However, their large parameter sizes present deployment challenges in resource-constrained environments with limited GPU memory capacity, as GPU memory is often insufficient to accommodate the full set of model weights. Consequently, typical deployments rely on CPU-GPU hybrid execution: the GPU handles compute-intensive GEMM operations, while the CPU processes the relatively lightweight attention mechanism. This setup introduces a key challenge: how to effectively optimize resource utilization across CPU and GPU? Prior work has designed system optimizations based on performance models with limited scope. Specifically, such models do not capture the complex interactions between hardware properties and system execution mechanisms. Therefore, previous approaches neither identify nor achieve the hardware limit. This paper presents MoE-Lens, a high-throughput MoE LLM inference system designed through holistic performance modeling for resource-constrained environments. Our performance model thoroughly analyzes various fundamental system components, including CPU memory capacity, GPU compute power, and workload characteristics, to understand the theoretical performance upper bound of MoE inference. Furthermore, it captures the system execution mechanisms to identify the key hardware bottlenecks and accurately predict the achievable throughput. Informed by our performance model, MoE-Lens introduces an inference system approaching hardware limits. Evaluated on diverse MoE models and datasets, MoE-Lens outperforms the state-of-the-art solution by 4.6x on average (up to 25.5x), with our theoretical model predicting performance with an average 94% accuracy.

cs.DC↗

arXiv2025

MoDM: Efficient Serving for Image Generation via Mixture-of-Diffusion Models

Diffusion-based text-to-image generation models trade latency for quality: small models are fast but generate lower-quality images, while large models produce better images but are slow. We present MoDM, a novel caching-based serving system for diffusion models that dynamically balances latency and quality through a mixture of diffusion models. Unlike prior approaches that rely on model-specific internal features, MoDM caches final images, allowing seamless retrieval and reuse across multiple diffusion model families. This design enables adaptive serving by dynamically balancing latency and image quality: using smaller models for cache-hit requests to reduce latency while reserving larger models for cache-miss requests to maintain quality. Small model image quality is preserved using retrieved cached images. We design a global monitor that optimally allocates GPU resources and balances inference workload, ensuring high throughput while meeting service-level objectives under varying request rates. Our evaluations show that MoDM significantly reduces average serving time by 2.5x while retaining image quality, making it a practical solution for scalable and resource-efficient model deployment.

cs.DC↗

arXiv2025

Vortex: Overcoming Memory Capacity Limitations in GPU-Accelerated Large-Scale Data Analytics

Despite the high computational throughput of GPUs, limited memory capacity and bandwidth-limited CPU-GPU communication via PCIe links remain significant bottlenecks for accelerating large-scale data analytics workloads. This paper introduces Vortex, a GPU-accelerated framework designed for data analytics workloads that exceed GPU memory capacity. A key aspect of our framework is an optimized IO primitive that leverages all available PCIe links in multi-GPU systems for the IO demand of a single target GPU. It routes data through other GPUs to such target GPU that handles IO-intensive analytics tasks. This approach is advantageous when other GPUs are occupied with compute-bound workloads, such as popular AI applications that typically underutilize IO resources. We also introduce a novel programming model that separates GPU kernel development from IO scheduling, reducing programmer burden and enabling GPU code reuse. Additionally, we present the design of certain important query operators and discuss a late materialization technique based on GPU's zero-copy memory access. Without caching any data in GPU memory, Vortex improves the performance of the state-of-the-art GPU baseline, Proteus, by 5.7$\times$ on average and enhances price performance by 2.5$\times$ compared to a CPU-based DuckDB baseline.

cs.DB↗

arXiv2025

Multi-Dimensional Vector ISA Extension for Mobile In-Cache Computing

In-cache computing technology transforms existing caches into long-vector compute units and offers low-cost alternatives to building expensive vector engines for mobile CPUs. Unfortunately, existing long-vector Instruction Set Architecture (ISA) extensions, such as RISC-V Vector Extension (RVV) and Arm Scalable Vector Extension (SVE), provide only one-dimensional strided and random memory accesses. While this is sufficient for typical vector engines, it fails to effectively utilize the large Single Instruction, Multiple Data (SIMD) widths of in-cache vector engines. This is because mobile data-parallel kernels expose limited parallelism across a single dimension. Based on our analysis of mobile vector kernels, we introduce a long-vector Multi-dimensional Vector ISA Extension (MVE) for mobile in-cache computing. MVE achieves high SIMD resource utilization and enables flexible programming by abstracting cache geometry and data layout. The proposed ISA features multi-dimensional strided and random memory accesses and efficient dimension-level masked execution to encode parallelism across multiple dimensions. Using a wide range of data-parallel mobile workloads, we demonstrate that MVE offers significant performance and energy reduction benefits of 2.9x and 8.8x, on average, compared to the SIMD units of a commercial mobile processor, at an area overhead of 3.6%.

cs.AR↗

arXiv2024

Palermo: Improving the Performance of Oblivious Memory using Protocol-Hardware Co-Design

Oblivious RAM (ORAM) hides the memory access patterns, enhancing data privacy by preventing attackers from discovering sensitive information based on the sequence of memory accesses. The performance of ORAM is often limited by its inherent trade-off between security and efficiency, as concealing memory access patterns imposes significant computational and memory overhead. While prior works focus on improving the ORAM performance by prefetching and eliminating ORAM requests, we find that their performance is very sensitive to workload locality behavior and incurs additional management overhead caused by the ORAM stash pressure. This paper presents Palermo: a protocol-hardware co-design to improve ORAM performance. The key observation in Palermo is that classical ORAM protocols enforce restrictive dependencies between memory operations that result in low memory bandwidth utilization. Palermo introduces a new protocol that overlaps large portions of memory operations, within a single and between multiple ORAM requests, without breaking correctness and security guarantees. Subsequently, we propose an ORAM controller architecture that executes the proposed protocol to service ORAM requests. The hardware is responsible for concurrently issuing memory requests as well as imposing the necessary dependencies to ensure a consistent view of the ORAM tree across requests. Using a rich workload mix, we demonstrate that Palermo outperforms the RingORAM baseline by 2.8x, on average, incurring a negligible area overhead of 5.78mm^2 (less than 2% in 12th generation Intel CPU after technology scaling) and 2.14W without sacrificing security. We further show that Palermo also outperforms the state-of-the-art works PageORAM, PrORAM, and IR-ORAM.

cs.CR↗

arXiv2024

Accurate and Fast Estimation of Temporal Motifs using Path Sampling

Counting the number of small subgraphs, called motifs, is a fundamental problem in social network analysis and graph mining. Many real-world networks are directed and temporal, where edges have timestamps. Motif counting in directed, temporal graphs is especially challenging because there are a plethora of different kinds of patterns. Temporal motif counts reveal much richer information and there is a need for scalable algorithms for motif counting. A major challenge in counting is that there can be trillions of temporal motif matches even with a graph with only millions of vertices. Both the motifs and the input graphs can have multiple edges between two vertices, leading to a combinatorial explosion problem. Counting temporal motifs involving just four vertices is not feasible with current state-of-the-art algorithms. We design an algorithm, TEACUPS, that addresses this problem using a novel technique of temporal path sampling. We combine a path sampling method with carefully designed temporal data structures, to propose an efficient approximate algorithm for temporal motif counting. TEACUPS is an unbiased estimator with provable concentration behavior, which can be used to bound the estimation error. For a Bitcoin graph with hundreds of millions of edges, TEACUPS runs in less than 1 minute, while the exact counting algorithm takes more than a day. We empirically demonstrate the accuracy of TEACUPS on large datasets, showing an average of 30$\times$ speedup (up to 2000$\times$ speedup) compared to existing GPU-based exact counting methods while preserving high count estimation accuracy.

cs.SI↗

arXiv2024

Understanding the Performance and Estimating the Cost of LLM Fine-Tuning

Due to the cost-prohibitive nature of training Large Language Models (LLMs), fine-tuning has emerged as an attractive alternative for specializing LLMs for specific tasks using limited compute resources in a cost-effective manner. In this paper, we characterize sparse Mixture of Experts (MoE) based LLM fine-tuning to understand their accuracy and runtime performance on a single GPU. Our evaluation provides unique insights into the training efficacy of sparse and dense versions of MoE models, as well as their runtime characteristics, including maximum batch size, execution time breakdown, end-to-end throughput, GPU hardware utilization, and load distribution. Our study identifies the optimization of the MoE layer as crucial for further improving the performance of LLM fine-tuning. Using our profiling results, we also develop and validate an analytical model to estimate the cost of LLM fine-tuning on the cloud. This model, based on parameters of the model and GPU architecture, estimates LLM throughput and the cost of training, aiding practitioners in industry and academia to budget the cost of fine-tuning a specific model.

cs.CL↗