MLManisha Luthra

Manisha Luthra

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Manisha Luthra has a total of 25 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    EcoLLM: Energy-Aware Benchmarking of LLMs for Data Processing Workloads

    by Pratyush Agnihotri, Manisha Luthra, Carsten Binnig

    Large language models (LLMs) are increasingly integrated into data management systems, yet their energy consumption remain largely unexplored. In this paper, we present EcoLLM, a workload-centric benchmark for studying energy-aware trade-offs of LLMs in data-processing tasks. EcoLLM models workload families, operator complexity, and data scale as well as supports both local and API-based models, and employs a hybrid energy measurement methodology. Our evaluation reveals several non-obvious findings. First, energy efficiency and task effectiveness are not aligned: highly efficient models can fail completely on pipeline generation tasks, while more energy-intensive models achieve correct results. Second, we observe a consistent trade-off between latency and energy, where lower latency is often achieved at disproportionately higher energy cost, leading to distinct execution regimes across deployment modes. Finally, per-task energy consumption appears small yet it scales to substantial cost at production workloads, making energy a critical factor in system design. These findings highlight the need for energy-aware and workload-aware model selection in LLM-based data systems.

  • Open Access

    Eighth International Workshop on Exploiting Artificial Intelligence Techniques for Data Management (aiDM)

    by Renata Borovica-Gajic, Manisha Luthra, Ryan Marcus, Rajesh Bordawekar, Oded Shmueli

  • Top Cited Papers
    Open Access

    How Good are Learned Cost Models, Really? Insights from Query Optimization Tasks

    by Roman Heinrich, Manisha Luthra, Johannes Wehrstein, Harald Kornmayer, Carsten Binnig

    Traditionally, query optimizers rely on cost models to choose the best execution plan from several candidates, making precise cost estimates critical for efficient query execution. In recent years, cost models based on machine learning have been proposed to overcome the weaknesses of traditional cost models. While these models have been shown to provide better prediction accuracy, only limited efforts have been made to investigate how well Learned Cost Models (LCMs) actually perform in query optimization and how they affect overall query performance. In this paper, we address this by a systematic study evaluating LCMs on three of the core query optimization tasks: join ordering, access path selection, and physical operator selection . In our study, we compare seven state-of-the-art LCMs to a traditional cost model and, surprisingly, find that the traditional model often still outperforms LCMs in these tasks. We conclude by highlighting major takeaways and recommendations to guide future research toward making LCMs more effective for query optimization.


  • Open Access

    How Good are Learned Cost Models, Really? Insights from Query Optimization Tasks

    by Roman Heinrich, Manisha Luthra, Johannes Wehrstein, Harald Kornmayer, Carsten Binnig

    Traditionally, query optimizers rely on cost models to choose the best execution plan from several candidates, making precise cost estimates critical for efficient query execution. In recent years, cost models based on machine learning have been proposed to overcome the weaknesses of traditional cost models. While these models have been shown to provide better prediction accuracy, only limited efforts have been made to investigate how well Learned Cost Models (LCMs) actually perform in query optimization and how they affect overall query performance. In this paper, we address this by a systematic study evaluating LCMs on three of the core query optimization tasks: join ordering, access path selection, and physical operator selection. In our study, we compare seven state-of-the-art LCMs to a traditional cost model and, surprisingly, find that the traditional model often still outperforms LCMs in these tasks. We conclude by highlighting major takeaways and recommendations to guide future research toward making LCMs more effective for query optimization.

  • Open Access

    No One Size (PPM) Fits All: Towards Privacy in Stream Processing Systems

    by Mikhail Fomichev, Manisha Luthra, Maik Benndorf, Pratyush Agnihotri

    Stream processing systems (SPSs) have been designed to process data streams in real-time, allowing organizations to analyze and act upon data on-the-fly, as it is generated. However, handling sensitive or personal data in these multilayered SPSs that distribute resources across sensor, fog, and cloud layers raises privacy concerns, as the data may be subject to unauthorized access and attacks that can violate user privacy, hence facing regulations such as the GDPR across the SPS layers. To address these issues, different privacy-preserving mechanisms (PPMs) are proposed to protect user privacy in SPSs. Yet, selecting and applying such PPMs in SPSs is challenging, since they must operate in real-time while tolerating little overhead. The multilayered nature of SPSs complicates privacy protection because each layer may confront different privacy threats, which must be addressed by specific PPMs. To overcome these challenges, we present Prinseps, our comprehensive privacy vision for SPSs. Towards this vision, we (1) identify critical privacy threats on different layers of the multilayered SPS, (2) evaluate the effectiveness of existing PPMs in addressing such threats, and (3) integrate privacy considerations into the decision-making processes of SPSs.


  • Open Access

    Zero-Shot Cost Models for Parallel Stream Processing

    by Pratyush Agnihotri, Boris Koldehofe, Carsten Binnig, Manisha Luthra

    This paper addresses the challenge of predicting the level of parallelism in distributed stream processing (DSP) systems, which are essential to deal with different high workload requirements of various industries such as e-commerce, online gaming, etc., where DSP systems are extensively used. Existing DSP systems rely on either manual tuning of parallelism degree or workload-driven learned models for tuning parallelism, which is either not efficient or can lead to costly operator migrations and downtime when there are workload drifts. Thus, we argue for a learned model that can autonomously decide on the right parallelism degree while generalizing across workloads and meeting the current demands of DSP applications. We propose a novel approach that leverages zero-shot cost models to predict parallelism degree while generalizing across unseen streaming workloads out-of-the-box. To reduce training effort, we propose a rule-based strategy that selects parallelism degree and meaningful transferable features related to query workload and hardware that influences the parallelism decisions. We demonstrate the effectiveness of our strategy by evaluating it with different amount of training queries and show that it achieves lower costs for parallel continuous query processing.


  • Open Access

    Distributed GPU Joins on Fast RDMA-capable Networks

    by Lasse Thostrup, Gloria Doci, Nils Boeschen, Manisha Luthra, Carsten Binnig

    In this paper, we present a novel pipelined GPU join that accelerates the performance of distributed DBMSs by leveraging GPU resources on fast networks. A key insight is that we enable pipelined join execution by overlapping the network shuffling with the build and probe phases, thereby significantly reducing the GPU idle time. To demonstrate this, we propose novel algorithms for distributed pipelined GPU joins with RDMA and GPUDirect for both arbitrarily large probe- and build-side tables. In our evaluation, we show our pipelined distributed GPU join can reduce the overall runtime of a full query by up to 6× against a state-of-the-art CPU-only join.


  • Open Access

    Efficient Complex Event Processing in Information-centric Networking at the Edge

    by Manisha Luthra, Johannes Pfannmüller, Boris Koldehofe, Jonas Höchst, Artur Sterz, Rhaban Hark, Bernd Freisleben

    Information-centric Networking (ICN) is an emerging Internet architecture that offers promising features, such as in-network caching and named data addressing, to support the edge computing paradigm, in particular Internet-of-Things (IoT) applications. ICN can benefit from Complex Event Processing (CEP), which is an in-network processing paradigm to specify and perform efficient query operations on data streams. However, integrating CEP into ICN is a challenging task due to the following reasons: (1) typical ICN architectures do not provide support for forwarding and processing continuous data streams; (2) IoT applications often need short response times and require robust event detection, which both are hard to accomplish using existing CEP systems. In this article, we present a novel network architecture, called INetCEP, for efficient CEP-based in-network processing as part of ICN. INetCEP enables efficient data processing in ICN by means of (1) a unified communication model that supports continuous data streams, (2) a meta query language for CEP to specify data processing operations in the data plane, and (3) query processing algorithms to resolve the specified operations. Our experimental results for two IoT use cases and datasets show that INetCEP offers very short response times of up to 73 {\mu}s under high workload and is more than 15X faster in terms of forwarding events than the state-of-the-art CEP system Flink. Furthermore, the delivery and processing of complex queries is around 32X faster than Flink and more than 100X faster than a naive pull-based reference approach, while maintaining 100% accuracy.