Serverless GPU Architecture for Enterprise HR Analytics: A Production-Scale BDaaS Implementation
Guilin Zhang 1, Wulan Guo 1, Ziqi Tan 1, Srinivas Vippagunta 3, Suchitra Raman 3, Shreeshankar Chatterjee 3 · 2025
Read Serverless GPU Architecture for Enterprise HR Analytics: A Production-Scale BDaaS Implementation by Guilin Zhang 1, Wulan Guo 1, Ziqi Tan 1, Srinivas Vippagunta 3, Suchitra Raman 3, Shreeshankar Chatterjee 3 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-19689.txt.
Introduction
Guilin Zhang and ten co-authors lay out a production blueprint for running enterprise human-resources analytics on serverless graphics processors. Large firms and agencies need workforce analytics that are fast, cheap, and auditable, while classic distributed frameworks like Spark add coordination and compliance overhead for moderate workloads. Their design pairs a single-node serverless GPU runtime with TabNet, an interpretable model for tabular data.
Benchmarks on human-resources and public datasets showed up to about four and a half times the throughput of Spark baselines, far lower latency, and about ninety percent lower cost per thousand inferences. Compliance mechanisms added only a few milliseconds, and the model's feature-mask explanations stayed stable under peak load, keeping the system auditable for regulated settings.
Catalog record
| CREATOR | Guilin Zhang 1, Wulan Guo 1, Ziqi Tan 1, Srinivas Vippagunta 3, Suchitra Raman 3, Shreeshankar Chatterjee 3 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 44 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-19689
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Serverless GPU Architecture for Enterprise HR Analytics: A Production-Scale BDaaS Implementation Thanks: Thanks: Corresponding author 4 : Hailong Jiang, [email protected] Guilin Zhang 1, Wulan Guo 1, Ziqi Tan 1, Srinivas Vippagunta 3, Suchitra Raman 3, Shreeshankar Chatterjee 3 Abstract Industrial and government organizations increasingly depend on data-driven analytics for workforce, finance, and regulated decision processes, where timeliness, cost efficiency, and compliance are critical. Distributed frameworks such as Spark and Flink remain effective for massive-scale batch or streaming analytics but introduce coordination complexity and auditing overheads that misalign with moderate-scale, latency-sensitive inference. Meanwhile, cloud providers now offer serverless GPUs, and models such as TabNet enable interpretable tabular ML, motivating new deployment blueprints for regulated environments. In this paper, we present a production-oriented Big Data as a Service (BDaaS) blueprint that integrates a single-node serverless GPU runtime with TabNet. The design leverages GPU acceleration for throughput, serverless elasticity for cost reduction, and feature-mask interpretability for IL4/FIPS compliance. We conduct benchmarks on the HR, Adult, and BLS datasets, comparing our approach against Spark and CPU baselines. Our results show that GPU pipelines achieve up to 4.5 × 4.5\times higher throughput, 98 × 98\times lower latency, and 90% lower cost per 1K inferences compared to Spark baselines, while compliance mechanisms add only ∼ \sim 5.7,ms latency with p 99 < p_{99}< 22,ms. Interpretability remains stable under peak load, ensuring reliable auditability. Taken together, these findings provide a compliance-aware benchmark, a reproducible Helm-packaged blueprint, and a decision framework that demonstrate the practicality of secure, interpretable, and cost-efficient serverless GPU analytics for regulated enterprise and government settings. I Introduction Industrial and government organizations increasingly rely on data-driven decision systems such as human resource (HR) analytics, financial risk forecasting, and other regulated processes to optimize operations and strengthen accountability [ 10 , 22 ] . Recent reviews highlight the rapid growth of people analytics and emphasize governance needs such as timeliness and auditability [ 5 , 27 ] . At the same time, enterprises are pushing for near-real-time insights rather than delayed batch reports, creating concrete requirements for low latency, cost efficiency, and compliance in production analytics stacks [ 28 ] . Distributed batch and micro-batch frameworks such as Hadoop/MapReduce and Apache Spark remain highly effective for large-scale ETL and historical analytics, but they often misalign with latency-sensitive inference at moderate scale. Spark’s micro-batch design, even with continuous mode, introduces coordination overhead that typically yields second- to sub-second latencies, which can be restrictive for interactive enterprise workloads [ 2 , 12 ] . Operating multi-node clusters further adds scheduling and fault-handling complexity, inflating both cost and compliance effort [ 17 ] . Comparative benchmarks confirm that Spark can exhibit higher event latency than Flink under recovery conditions [ 18 ] , and while Flink offers event-time processing and exactly-once semantics at terabyte scale with RocksDB backends [ 8 ] , its distributed runtime requires tuning, monitoring, and auditing overhead that can be disproportionate for moderate-size inference workloads [ 32 ] . In parallel, cloud providers have introduced serverless GPU offerings with per-second billing, scale-to-zero, and autoscaling, while research prototypes such as SAGE and Torpor demonstrate system designs that reduce cold-start costs and improve GPU sharing for low-latency inference [ 14 , 23 , 13 , 34 , 31 ] . Together with the well-established throughput and latency advantages of GPUs for ML inference, these trends suggest that a single-node serverless GPU architecture can replace multi-node CPU clusters for many moderate-scale, latency-critical workloads, reducing operational burden and cost. Regulated domains such as public sector, healthcare, and defense introduce additional requirements. Deployments must comply with FIPS 140-3 and DoD IL4/CC SRG controls, where cryptographic enforcement (mTLS, JWT, auditing) adds measurable latency overhead [ 7 , 35 ] . Any candidate architecture must quantify this security budget while still meeting p_{95} / p_{99} SLAs. At the same time, interpretable ML is increasingly mandated for regulated workforce decisions; TabNet provides step-wise feature-mask attributions for tabular data, making it a strong fit for HR analytics where case-level explanations support audits and monitoring [ 3 ] . Together, these limitations call for a new architectural approach that explicitly balances performance, compliance, and interpretability in industrial and government environments. In this paper, we introduce a production-oriented Big Data as a Service (BDaaS) blueprint that integrates a single-node serverless GPU runtime with TabNet for tabular ML inference. The blueprint is designed around three pillars: (i) GPU acceleration to achieve high-throughput, low-latency inference for moderate-scale workloads; (ii) serverless elasticity and pay-per-use economics to simplify operations and reduce cost compared to distributed clusters; and (iii) built-in interpretability via TabNet’s feature masks, enabling audit-ready explanations under IL4/FIPS compliance constraints. This architecture provides a streamlined alternative to multi-node Spark or Flink deployments, reducing system complexity while satisfying the stringent requirements of regulated enterprise analytics. We evaluate this blueprint through three research questions: • RQ-1: How does a single-node GPU serverless pipeline compare with distributed Spark/CPU clusters in terms of throughput, latency, and cost for moderate - scale workloads ? • RQ-2: What is the measurable overhead introduced by IL4/FIPS compliance controls (mTLS, JWT, auditing), and can SLA targets (e.g., p_{95}/p_{99} latency) still be met? • RQ-3: Do TabNet’s feature-mask explanations remain stable under production-scale, high-throughput inference, ensuring interpretability for regulated analytics? To investigate these questions, we conduct controlled experiments using three public datasets (HR, Adult, and BLS), benchmarking our serverless GPU–TabNet blueprint against multi-node Spark and CPU baselines. We systematically measure throughput, latency, and cost under varying batch sizes, quantify the latency overhead of security mechanisms, and evaluate the stability of interpretability under high-throughput workloads. Our evaluation indicates that: (i) GPU serverless pipelines significantly outperform Spark clusters at batch sizes \geq 200, achieving up to 4.5\times higher throughput and 98\times lower latency at \text{batch}=1000 , with cost reductions of up to 90% per 1K inferences (RQ-1); (ii) IL4/FIPS compliance mechanisms introduce only \sim 5.7 ms additional latency, preserving p_{99}<20 ms and showing that strict security can coexist with low-latency inference (RQ-2); and (iii) TabNet feature-mask stability remains above 0.88 even under peak throughput, ensuring consistent, audit-ready interpretability in production settings (RQ-3). We further discuss deployment trade-offs, limitations, and potential optimizations in Section VI . Our main contributions in this work are: 1. Compliance-aware benchmarking: To the best of our knowledge, this is the first work to conduct a systematic component-level decomposition of IL4/FIPS overhead for serverless ML inference, quantifying a 5.7 ms security budget under steady-state operation. 2. Industrial decision framework: We provide actionable guidance on CPU vs GPU vs Spark trade-offs, including batch-size thresholds where GPU serverless becomes cost-effective and when to prefer distributed frameworks. 3. Interpretability at scale: We demonstrate that TabNet’s feature-mask stability ( > 0.88) persists under high-throughput inference, ensuring audit-ready explanations for sensitive HR analytics. 4. Reproducible blueprint: We release a complete Helm-packaged serverless GPU deployment with integrated security mesh and monitoring, enabling practitioners in industry and government to adopt our approach with a streamlined process. The rest of this paper is organized as follows. Section II introduces the related work of this paper. Section III introduces the system architecture of our compliance-aware serverless GPU blueprint. Section IV describes the experimental setup, including datasets, model configuration, baselines, and evaluation metrics. Section V presents the