arXiv subjects

Chandra Narayanaswami

Publications and source records attributed to Chandra Narayanaswami.

7 recordsLinked to original sources

arXiv2025

Story of Two GPUs: Characterizing the Resilience of Hopper H100 and Ampere A100 GPUs

This study characterizes GPU resilience in Delta, a large-scale AI system that consists of 1,056 A100 and H100 GPUs, with over 1,300 petaflops of peak throughput. We used 2.5 years of operational data (11.7 million GPU hours) on GPU errors. Our major findings include: (i) H100 GPU memory resilience is worse than A100 GPU memory, with 3.2x lower per-GPU MTBE for memory errors, (ii) The GPU memory error-recovery mechanisms on H100 GPUs are insufficient to handle the increased memory capacity, (iii) H100 GPUs demonstrate significantly improved GPU hardware resilience over A100 GPUs with respect to critical hardware components, (iv) GPU errors on both A100 and H100 GPUs frequently result in job failures due to the lack of robust recovery mechanisms at the application level, and (v) We project the impact of GPU node availability on larger-scales and find that significant overprovisioning of 5% is necessary to handle GPU failures.

cs.DC↗

arXiv2025

CPU-Limits kill Performance: Time to rethink Resource Control

Research in compute resource management for cloud-native applications is dominated by the problem of setting optimal CPU limits -- a fundamental OS mechanism that strictly restricts a container's CPU usage to its specified CPU-limits . Rightsizing and autoscaling works have innovated on allocation/scaling policies assuming the ubiquity and necessity of CPU-limits . We question this. Practical experiences of cloud users indicate that CPU-limits harms application performance and costs more than it helps. These observations are in contradiction to the conventional wisdom presented in both academic research and industry best practices. We argue that this indiscriminate adoption of CPU-limits is driven by erroneous beliefs that CPU-limits is essential for operational and safety purposes. We provide empirical evidence making a case for eschewing CPU-limits completely from latency-sensitive applications. This prompts a fundamental rethinking of auto-scaling and billing paradigms and opens new research avenues. Finally, we highlight specific scenarios where CPU-limits can be beneficial if used in a well-reasoned way (e.g. background jobs).

cs.DC↗

arXiv2025

An Optimistic-Robust Approach for Dynamic Positioning of Omnichannel Inventories

We introduce a new class of data-driven and distribution-free optimistic-robust bimodal inventory optimization (BIO) strategy to effectively allocate inventory across a retail chain to meet time-varying, uncertain omnichannel demand. The bimodal nature of BIO stems from its ability to balance downside risk, as in traditional Robust Optimization (RO), which focuses on worst-case adversarial demand, with upside potential to enhance average-case performance. This enables BIO to remain as resilient as RO while capturing benefits that would otherwise be lost due to endogenous outliers. Omnichannel inventory planning provides a suitable problem setting for analyzing the effectiveness of BIO's bimodal strategy in managing the tradeoff between lost sales at stores and cross-channel e-commerce fulfillment costs, factors that are inherently asymmetric due to channel-specific behaviors. We provide structural insights about the BIO solution and how it can be tuned to achieve a preferred tradeoff between robustness and the average-case performance. Using a real-world dataset from a large American omnichannel retail chain, a business value assessment during a peak period indicates that BIO outperforms pure RO by 27% in terms of realized average profitability and surpasses other competitive baselines under imperfect distributional information by over 10%. This demonstrates that BIO provides a novel, data-driven, and distribution-free alternative to traditional RO that achieves strong average performance while carefully balancing robustness.

math.OC↗

arXiv2025

Queue management for slo-oriented large language model serving

Large language model (LLM) serving is becoming an increasingly critical workload for cloud providers. Existing LLM serving systems focus on interactive requests, such as chatbots and coding assistants, with tight latency SLO requirements. However, when such systems execute batch requests that have relaxed SLOs along with interactive requests, it leads to poor multiplexing and inefficient resource utilization. To address these challenges, we propose QLM, a queue management system for LLM serving. QLM maintains batch and interactive requests across different models and SLOs in a request queue. Optimal ordering of the request queue is critical to maintain SLOs while ensuring high resource utilization. To generate this optimal ordering, QLM uses a Request Waiting Time (RWT) Estimator that estimates the waiting times for requests in the request queue. These estimates are used by a global scheduler to orchestrate LLM Serving Operations (LSOs) such as request pulling, request eviction, load balancing, and model swapping. Evaluation on heterogeneous GPU devices and models with real-world LLM serving dataset shows that QLM improves SLO attainment by 40-90% and throughput by 20-400% while maintaining or improving device utilization compared to other state-of-the-art LLM serving systems. QLM's evaluation is based on the production requirements of a cloud provider. QLM is publicly available at https://www.github.com/QLM-project/QLM.

cs.DC↗

arXiv2025

Hierarchical Autoscaling for Large Language Model Serving with Chiron

Large language model (LLM) serving is becoming an increasingly important workload for cloud providers. Based on performance SLO requirements, LLM inference requests can be divided into (a) interactive requests that have tight SLOs in the order of seconds, and (b) batch requests that have relaxed SLO in the order of minutes to hours. These SLOs can degrade based on the arrival rates, multiplexing, and configuration parameters, thus necessitating the use of resource autoscaling on serving instances and their batch sizes. However, previous autoscalers for LLM serving do not consider request SLOs leading to unnecessary scaling and resource under-utilization. To address these limitations, we introduce Chiron, an autoscaler that uses the idea of hierarchical backpressure estimated using queue size, utilization, and SLOs. Our experiments show that Chiron achieves up to 90% higher SLO attainment and improves GPU efficiency by up to 70% compared to existing solutions.

cs.DC↗

arXiv2024

Optimal Workload Placement on Multi-Instance GPUs

There is an urgent and pressing need to optimize usage of Graphical Processing Units (GPUs), which have arguably become one of the most expensive and sought after IT resources. To help with this goal, several of the current generation of GPUs support a partitioning feature, called Multi-Instance GPU (MIG) to allow multiple workloads to share a GPU, albeit with some constraints. In this paper we investigate how to optimize the placement of Large Language Model (LLM)-based AI Inferencing workloads on GPUs. We first identify and present several use cases that are encountered in practice that require workloads to be efficiently placed or migrated to other GPUs to make room for incoming workloads. The overarching goal is to use as few GPUs as possible and to further minimize memory and compute wastage on GPUs that are utilized. We have developed two approaches to address this problem: an optimization method and a heuristic method. We benchmark these with two workload scheduling heuristics for multiple use cases. Our results show up to 2.85x improvement in the number of GPUs used and up to 70% reduction in GPU wastage over baseline heuristics. We plan to enable the SRE community to leverage our proposed method in production environments.

cs.DC↗

arXiv2023

Hierarchical Proxy Modeling for Improved HPO in Time Series Forecasting

Selecting the right set of hyperparameters is crucial in time series forecasting. The classical temporal cross-validation framework for hyperparameter optimization (HPO) often leads to poor test performance because of a possible mismatch between validation and test periods. To address this test-validation mismatch, we propose a novel technique, H-Pro to drive HPO via test proxies by exploiting data hierarchies often associated with time series datasets. Since higher-level aggregated time series often show less irregularity and better predictability as compared to the lowest-level time series which can be sparse and intermittent, we optimize the hyperparameters of the lowest-level base-forecaster by leveraging the proxy forecasts for the test period generated from the forecasters at higher levels. H-Pro can be applied on any off-the-shelf machine learning model to perform HPO. We validate the efficacy of our technique with extensive empirical evaluation on five publicly available hierarchical forecasting datasets. Our approach outperforms existing state-of-the-art methods in Tourism, Wiki, and Traffic datasets, and achieves competitive result in Tourism-L dataset, without any model-specific enhancements. Moreover, our method outperforms the winning method of the M5 forecast accuracy competition.

cs.LG↗