arXiv subjects

Partha Kanuparthy

Publications and source records attributed to Partha Kanuparthy.

8 recordsLinked to original sources

arXiv2026

TLX: Hardware-Native, Evolvable MIMW GPU Compiler for Large-scale Production Environments

Modern GPUs increasingly rely on specialized hardware units and asynchronous coordination mechanisms, so performance depends on orchestrating data movement, tensor-core computation, and synchronization rather than exposing more thread-level parallelism. This creates a programming-model tension: if too much execution structure is hidden, the compiler must catch up to new hardware mechanisms; if too much is exposed, the burden of orchestration falls back onto the programmer. We present TLX (Triton Low-level Language Extensions), built around MIMW (Multi-Instruction, Multi-Warp), which expresses orchestration at warp-group granularity while preserving Triton's productive blocked programming model for regular computation. TLX realizes this idea as an embedded extension to Triton, exposing explicit interfaces for multi-warp execution, local-memory orchestration, asynchronous operations, and cluster-aware control. Our evaluation shows that TLX supports substantial customization with limited development effort while remaining competitive with state-of-the-art implementations. TLX-authored kernels have been deployed in large-scale training and inference production systems. Our code is open sourced at https://github.com/facebookexperimental/triton.

cs.AR↗

arXiv2025

TorchTraceAP: A New Benchmark Dataset for Detecting Performance Anti-Patterns in Computer Vision Models

Identifying and addressing performance anti-patterns in machine learning (ML) models is critical for efficient training and inference, but it typically demands deep expertise spanning system infrastructure, ML models and kernel development. While large tech companies rely on dedicated ML infrastructure engineers to analyze torch traces and benchmarks, such resource-intensive workflows are largely inaccessible to computer vision researchers in general. Among the challenges, pinpointing problematic trace segments within lengthy execution traces remains the most time-consuming task, and is difficult to automate with current ML models, including LLMs. In this work, we present the first benchmark dataset specifically designed to evaluate and improve ML models' ability to detect anti patterns in traces. Our dataset contains over 600 PyTorch traces from diverse computer vision models classification, detection, segmentation, and generation collected across multiple hardware platforms. We also propose a novel iterative approach: a lightweight ML model first detects trace segments with anti patterns, followed by a large language model (LLM) for fine grained classification and targeted feedback. Experimental results demonstrate that our method significantly outperforms unsupervised clustering and rule based statistical techniques for detecting anti pattern regions. Our method also effectively compensates LLM's limited context length and reasoning inefficiencies.

cs.CV↗

arXiv2025

TritonForge: Profiling-Guided Framework for Automated Triton Kernel Optimization

High-performance GPU kernel optimization remains a critical yet labor-intensive task in modern machine learning workloads. Although Triton, a domain-specific language for GPU programming, enables developers to write efficient kernels with concise code, achieving expert-level performance still requires deep understanding of GPU architectures and low-level performance trade-offs. We present TritonForge, a profiling-guided framework for automated Triton kernel optimization. TritonForge integrates kernel analysis, runtime profiling, and iterative code transformation to streamline the optimization process. By incorporating feedback from profiling results, the system identifies performance bottlenecks, proposes targeted code modifications, and evaluates their impact automatically. Across diverse kernel types, TritonForge achieves up to 5x performance improvement over baseline implementations and on average 1.76x of the cases are successful, providing a foundation for future research in automated GPU performance optimization.

cs.SE↗

arXiv2016

YTrace: End-to-end Performance Diagnosis in Large Cloud and Content Providers

Content providers build serving stacks to deliver content to users. An important goal of a content provider is to ensure good user experience, since user experience has an impact on revenue. In this paper, we describe a system at Yahoo called YTrace that diagnoses bad user experience in near real time. We present the different components of YTrace for end-to-end multi-layer diagnosis (instrumentation, methods and backend system), and the system architecture for delivering diagnosis in near real time across all user sessions at Yahoo. YTrace diagnoses problems across service and network layers in the end-to-end path spanning user host, Internet, CDN and the datacenters, and has three diagnosis goals: detection, localization and root cause analysis (including cascading problems) of performance problems in user sessions with the cloud. The key component of the methods in YTrace is capturing and discovering causality, which we design based on a mix of instrumentation API, domain knowledge and blackbox methods. We show three case studies from production that span a large-scale distributed storage system, a datacenter-wide network, and an end-to-end video serving stack at Yahoo. We end by listing a number of open directions for performance diagnosis in cloud and content providers.

cs.DC↗

arXiv2016

Finding Needles in the Haystack: Harnessing Syslogs for Data Center Management

Network device syslogs are ubiquitous and abundant in modern data centers with most large data centers producing millions of messages per day. Yet, the operational information reflected in syslogs and their implications on diagnosis or management tasks are poorly understood. Prevalent approaches to understanding syslogs focus on simple correlation and abnormality detection and are often limited to detection providing little insight towards diagnosis and resolution. Towards improving data center operations, we propose and implement Log-Prophet, a system that applies a toolbox of statistical techniques and domain-specific models to mine detailed diagnoses. Log-Prophet infers causal relationships between syslog lines and constructs succinct but valuable problem graphs, summarizing root causes and their locality, including cascading problems. We validate Log-Prophet using problem tickets and through operator interviews. To demonstrate the strength of Log-Prophet, we perform an initial longitudinal study of a large online service provider's data center. Our study demonstrates that Log-Prophet significantly reduces the number of alerts while highlighting interesting operational issues.

cs.NI↗

arXiv2016

Performance Characterization of a Commercial Video Streaming Service

Despite the growing popularity of video streaming over the Internet, problems such as re-buffering and high startup latency continue to plague users. In this paper, we present an end-to-end characterization of Yahoo's video streaming service, analyzing over 500 million video chunks downloaded over a two-week period. We gain unique visibility into the causes of performance degradation by instrumenting both the CDN server and the client player at the chunk level, while also collecting frequent snapshots of TCP variables from the server network stack. We uncover a range of performance issues, including an asynchronous disk-read timer and cache misses at the server, high latency and latency variability in the network, and buffering delays and dropped frames at the client. Looking across chunks in the same session, or destined to the same IP prefix, we see how some performance problems are relatively persistent, depending on the video's popularity, the distance between the client and server, and the client's operating system, browser, and Flash runtime.

cs.NI↗

arXiv2012

DNS-based Ingress Load Balancing: An Experimental Evaluation

Multihomed services can load-balance their incoming connection requests using DNS, resolving the name of the server with different addresses depending on the link load that corresponds to each address. Previous work has studied a number of problems with this approach, e.g., due to Time-to-Live duration violations and client proximity to local DNS servers. In this paper, we experimentally evaluate a DNS-based ingress traffic engineering system that we deployed at Georgia Tech. Our objective is to understand whether simple and robust load balancing algorithms can be accurate in practice, despite aforementioned problems with DNS-based load balancing methods. In particular, we examine the impact of various system parameters and of the main workload characteristics. We show that a window-based measurement scheme can be fairly accurate in practice, as long as its window duration has been appropriately configured.

cs.NI↗

arXiv2011

Can User-Level Probing Detect and Diagnose Common Home-WLAN Pathologies?

Common WLAN pathologies include low signal-to-noise ratio, congestion, hidden terminals or interference from non-802.11 devices and phenomena. Prior work has focused on the detection and diagnosis of such problems using layer-2 information from 802.11 devices and special-purpose access points and monitors, which may not be generally available. Here, we investigate a userlevel approach: is it possible to detect and diagnose 802.11 pathologies with strictly user-level active probing, without any cooperation from, and without any visibility in, layer-2 devices? In this paper, we present preliminary but promising results indicating that such diagnostics are feasible.

cs.NI↗