IHIsraat Haque

Israat Haque

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Israat Haque has a total of 60 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Performance Characterization of dApps in Open Radio Access Networks

    by Conrado Boeira, Eduardo Baena, Andrea Lacava, Tommaso Melodia, Dimitrios Koutsonikolas, Israat Haque

    Despite recommendations to deploy real-time Open Radio Access Network (O-RAN) applications (dApps) in containerized environments, existing approaches predominantly rely on bare-metal servers. Moreover, current dApp deployments offer limited visibility into the resource usage patterns of both intelligent and non-intelligent dApps, hindering informed deployment decisions. This work addresses these gaps by implementing and evaluating representative dApps across multiple deployment scenarios (bare-metal and containers) to characterize the trade-offs in latency, scalability, and resource utilization. Additionally, we identify key performance bottlenecks and demonstrate how offloading dApps to emerging hardware accelerators, such as smart Network Interface Cards (NICs), can alleviate these limitations and improve real-time responsiveness in O-RAN systems.


  • Open Access

    Characterizing Performance–Energy Trade-offs of Large Language Models in Multi-Request Workflows

    by Md. Monzurul Amin Ifath, Israat Haque

    Large language models (LLMs) are increasingly deployed in applications forming multi-request workflows like document summarization, search-based copilots, and multi-agent programming. While these workflows unlock richer functionality, they also amplify latency and energy demand during inferences. Existing measurement and benchmarking efforts either focus on assessing performance of LLM inference systems or consider single-request evaluations, overlooking workflow dependencies and cross-request interactions unique to multi-request workflows. Moreover, the energy usage of such interdependent LLM calls is not explored in-depth. To address these gaps, this paper presents the first systematic characterization of performance--energy trade-offs in multi-request LLM inference. We develop and evaluate four representative workloads that capture sequential, interactive, agentic, and composite patterns common in modern deployments. Using an empirical NVIDIA A100 testbed with state-of-the-art serving systems (vLLM and Parrot), we systematically analyze how key energy knobs (e.g., input-output length, batch size, and GPU power cap) reshape latency, throughput, and component-level (e.g., CPU, GPU, and DRAM) energy use. Our findings reveal that batch size is the most impactful lever, though its benefits are highly workload dependent. While optimal batching benefits workloads with large shared prompts, it is ineffective for sequential summarization and only partially effective for mutli-agent coding. GPU power capping provides modest but predictable savings, while output length induces linear energy scaling with limited efficiency gains. We further demonstrate that engine-level optimizations in vLLM (e.g., continuous batching, PagedAttention) maintain higher GPU utilization and efficiency, especially for decode-heavy workloads, while Parrot's workflow-aware scheduling achieves lower energy consumption under stringent power constraints. These findings offer actionable guidelines for developers and system operators in designing performance- and energy-aware LLM serving systems in emerging multi-request workflows.

  • Open Access

    Characterizing Performance–Energy Trade-offs of Large Language Models in Multi-Request Workflows

    by Md. Monzurul Amin Ifath, Israat Haque

    Large language models (LLMs) are increasingly deployed in applications forming multi-request workflows like document summarization, search-based copilots, and multi-agent programming. While these workflows unlock richer functionality, they also amplify latency and energy demand during inferences. Existing measurement and benchmarking efforts either focus on assessing performance of LLM inference systems or consider single-request evaluations, overlooking workflow dependencies and cross-request interactions unique to multi-request workflows. Moreover, the energy usage of such interdependent LLM calls is not explored in-depth. To address these gaps, this paper presents the first systematic characterization of performance--energy trade-offs in multi-request LLM inference. We develop and evaluate four representative workloads that capture sequential, interactive, agentic, and composite patterns common in modern deployments. Using an empirical NVIDIA A100 testbed with state-of-the-art serving systems (vLLM and Parrot), we systematically analyze how key energy knobs (e.g., input-output length, batch size, and GPU power cap) reshape latency, throughput, and component-level (e.g., CPU, GPU, and DRAM) energy use. Our findings reveal that batch size is the most impactful lever, though its benefits are highly workload dependent. While optimal batching benefits workloads with large shared prompts, it is ineffective for sequential summarization and only partially effective for mutli-agent coding. GPU power capping provides modest but predictable savings, while output length induces linear energy scaling with limited efficiency gains. We further demonstrate that engine-level optimizations in vLLM (e.g., continuous batching, PagedAttention) maintain higher GPU utilization and efficiency, especially for decode-heavy workloads, while Parrot's workflow-aware scheduling achieves lower energy consumption under stringent power constraints. These findings offer actionable guidelines for developers and system operators in designing performance- and energy-aware LLM serving systems in emerging multi-request workflows.


  • Open Access

    Characterizing Performance–Energy Trade-offs of Large Language Models in Multi-Request Workflows

    by Monzurul Amin Ifath, Israat Haque

    Large language models (LLMs) are increasingly deployed in applications forming multi-request workflows like document summarization, search-based copilots, and multi-agent programming. While these workflows unlock richer functionality, they also amplify latency and energy demand during inferences. Existing measurement and benchmarking efforts either focus on assessing performance of LLM inference systems or consider single-request evaluations, overlooking workflow dependencies and cross-request interactions unique to multi-request workflows. Moreover, the energy usage of such interdependent LLM calls is not explored in-depth. To address these gaps, this paper presents the first systematic characterization of performance--energy trade-offs in multi-request LLM inference. We develop and evaluate four representative workloads that capture sequential, interactive, agentic, and composite patterns common in modern deployments. Using an empirical NVIDIA A100 testbed with state-of-the-art serving systems (vLLM and Parrot), we systematically analyze how key energy knobs (e.g., input-output length, batch size, and GPU power cap) reshape latency, throughput, and component-level (e.g., CPU, GPU, and DRAM) energy use. Our findings reveal that batch size is the most impactful lever, though its benefits are highly workload dependent. While optimal batching benefits workloads with large shared prompts, it is ineffective for sequential summarization and only partially effective for mutli-agent coding. GPU power capping provides modest but predictable savings, while output length induces linear energy scaling with limited efficiency gains. We further demonstrate that engine-level optimizations in vLLM (e.g., continuous batching, PagedAttention) maintain higher GPU utilization and efficiency, especially for decode-heavy workloads, while Parrot's workflow-aware scheduling achieves lower energy consumption under stringent power constraints. These findings offer actionable guidelines for developers and system operators in designing performance- and energy-aware LLM serving systems in emerging multi-request workflows.

  • Keep Calm But Log for Trouble: Coordination-Free Fault-Tolerance for In-Switch Applications

    by Ricardo Parizotto, Israat Haque, Alberto Schaeffer-Filho

    In-network computing (INC) offloads parts of the functionality of a distributed system to programmable switches. Once we move the computation into the network, failures may cause loss of essential information and disrupt the operation of these systems. To ensure high availability in the event of switch failures, the current state-of-the-art replicates state across multiple INCs. They achieve this by using state-machine replication, ensuring that INC replicas are consistent by coordinating the replication between multiple INC nodes. In this paper, we demonstrate that decoupling the consistency guarantees from the replication reduces the overhead of INC fault tolerance. We present RESIST, a system for building fault-tolerant INC using asynchronous replication and replay-based recovery. We propose new techniques for logging information and different replay techniques to restore INC systems according to consistency semantics. Furthermore, we apply RESIST techniques to existing INC functionalities, including event synchronization for distributed simulations and aggregation for distributed training. Our prototype of RESIST enables fault tolerance for INC applications on BMv2 and in a testbed with Tofino ASICs. Experiments show that the system provides fault tolerance with negligible overhead for non-failure scenarios and recovers from failures in less than 0.2 s.

  • A Unified Convergence Analysis of Decentralized Federated Learning at the Edge

    by Chengyan Jiang, Jiamin Fan, Talal Halabi, Israat Haque

    Federated learning (FL) reshaped the AI model training paradigm by incorporating privacy preserved approach of collaborative model training without sharing model data but only parameters. This learning can be centralized FL (CFL) or decentralized federated learning (DFL), where CFL may suffer from latency and robustness due to the reliance on a coordinating server for model convergence. On the other hand, DFL enables direct collaboration among participating devices without relying on a central server. Each device can independently connect to other devices and share model parameters. In such collaborative training paradigm, model convergence in the presence of various deployment topologies, AI model types, nonindependent identically distributed (non-i.i.d.) data distribution, and training strategies demands systematic analysis to realize their practical deployment in critical applications like intelligent transportation, smart factories, and real-time surveillance. A couple of works attempt to conduct only partial analysis and completely neglected incorporating non-i.i.d. data distribution, a critical factor in practical deployment of DFL in mentioned applications. This work conducts a systematic analysis on the convergence of DFL considering a wide range of AI models [e.g., classical, deep neural networks and large language models (LLMs)], network topologies (e.g., linear, ring, star, and mesh), training strategies (e.g., continuous and aggregate), and degree of non-i.i.d. data distributions. The analysis includes both mathematical formulations and their implementation and evaluation using real world data. The results confirm that the convergence rate of the models is inversely proportional to the degree of non-i.i.d. data distribution. Moreover, judicial selection of network topologies and training strategies can aid in this convergence process for the practical deployment of DFL on the edge.


  • Reinforcement Learning-Based In-Network Load Balancing

    by Hesam Tajbakhsh, Ricardo Parizotto, Alberto Schaeffer-Filho, Israat Haque

    Ensuring consistent performance becomes increasingly challenging with the growing complexity of applications in data centers. This is where load balancing emerges as a vital component. A load balancer distributes network or application traffic across various servers, resources, or pathways. In this article, we present P4WISE, a load balancer designed for software-defined networks. Operating on both the data and control planes, it employs reinforcement learning to distribute computational loads with granularity at inter and intra-server levels. Evaluation results demonstrate a remarkable 90% accuracy in predicting the optimal load balancing strategy of P4WISE in dynamic scenarios. Notably, unlike supervised or unsupervised methods, it eliminates the need for retraining when the environment undergoes minor or major changes. Instead, P4WISE autonomously adjusts and retrains itself based on observed states within the data center.


  • Open Access

    A Generalized GNN-Transformer-Based Radio Link Failure Prediction Framework in 5G RAN

    by Kazi Hasan, Khaleda Papry, Thomas Trappenberg, Israat Haque

    Radio Link Failure (RLF) prediction system in Radio Access Networks (RANs) is critical for ensuring seamless communication and meeting the stringent requirements of high data rates, low latency, and improved reliability in 5G networks. However, weather conditions such as precipitation, humidity, temperature, and wind impact these communication links. Usually, historical radio link Key Performance Indicators (KPIs) and their surrounding weather station observations are utilized for building learning-based RLF prediction models. However, such models must be capable of learning the spatial weather context in a dynamic RAN and effectively encoding time series KPIs with the weather observation data. Existing work utilizes a heuristic-based and non-generalizable weather station aggregation method that uses Long Short-Term Memory (LSTM) for non-weighted sequence modeling. This paper fills the gap by proposing GenTrap, a novel RLF prediction framework that introduces a Graph Neural Network (GNN)-based learnable weather effect aggregation module and employs state-of-the-art time series transformer as the temporal feature extractor for radio link failure prediction. The GNN module encodes surrounding weather station data of each radio site while the transformer module encodes historical radio and weather observation features. The proposed aggregation method of GenTrap can be integrated into any existing prediction model to achieve better performance and generalizability. We evaluate GenTrap on two real-world datasets (rural and urban) with 2.6 million KPI data points and show that GenTrap offers a significantly higher F1-score of 0.93 for rural and 0.79 for urban, an increase of 29% and 21% respectively, compared to the state-of-the-art LSTM-based solutions while offering a 20% increased generalization capability.