DBOS Network Sensing: A Web Services Approach to Collaborative Awareness

Sophia Lockton, Jeremy Kepner, Michael Stonebraker, Hayden Jananthan, LaToya Anderson, William Arcand, David Bestor, William Bergeron, Alex Bonn, Daniel Burrill, Chansup Byun, Timothy Davis, Vijay Gadepally, Michael Houle, Matthew Hubbell, Piotr Luszczek, Peter Michaleas, Lauren Milechin, Chasen Milner, Guillermo Morales, Julie Mullen, Michel Pelletier, Alex Poliakov, Andrew Prout, Albert Reuther, Antonio Rosa, Charles Yee, Alex Pentland

Abstract

Modern cybersecurity often faces a trade-off between isolated, ineffective monitoring and privacy-invasive data sharing. This paper presents a novel approach that integrates network sensing directly into the operating system layer. By leveraging the Database Operating System (DBOS) and GraphBLAS, we demonstrate a "collaborative awareness" framework. This system represents network traffic as hypersparse matrices, allowing distributed services to share anonymized patterns to detect coordinated threats. Benchmarks on the MIT SuperCloud show the system sustains over 100,000 requests per second with linear scalability, proving that high-performance security can be an intrinsic feature of cloud infrastructure.

Introduction

Traditional security architectures typically rely on Intrusion Detection Systems (IDSs) that operate in "stovepipes"—isolated silos that monitor only local traffic. While effective for local anomalies, these silos fail to detect sophisticated, coordinated attacks spanning multiple services. Conversely, "radical transparency," where organizations share raw logs, is often precluded by privacy regulations and data volume.

The spectrum of network awareness, positioning Collaborative Awareness as the ideal middle ground between isolated stovepipes and radical transparency.

We propose Collaborative Awareness as a middle ground. The objective is to build a system capable of aggregating and analyzing network traffic across distributed nodes in real-time, maintaining minimal computational overhead while preserving user privacy.

Methodology: DBOS and GraphBLAS

The core of our approach lies in the mathematical representation of network traffic. Instead of traditional text-based logging, we utilize hypersparse matrices via the GraphBLAS standard.

Hypersparse Representation

Network traffic is represented as source/destination IP pairs within a matrix of $2^{32} \times 2^{32}$ dimensions. Despite the massive theoretical size, the actual traffic is "hypersparse," allowing GraphBLAS to perform high-speed linear algebra operations to calculate metrics like unique links or fan-out with extreme efficiency.

System Integration

We implemented this sensing within the Database Operating System (DBOS). In DBOS, system services are transactional procedures on a high-performance database. This allows monitoring to be embedded directly into the "runtime fabric" of the OS. We explored two implementation paths: a high-throughput Python-GraphBLAS approach (using pPython) and a durable OneSparse PostgreSQL extension.

The four-step workflow for network sensing: 1) Extracting IP vectors from HTTP requests, 2) Constructing base traffic matrices, 3) Aggregating matrices locally, and 4) Computing network metrics.

Hierarchical Aggregation

As shown in the workflow, local worker nodes aggregate traffic into base matrices. These are periodically merged into a locally aggregated matrix, which is then sent to a coordinator node to form a global view of the network. This hierarchy ensures that only structural patterns—rather than raw user data—are shared.

Performance Evaluation

To evaluate the system, we used synthetic traffic based on the IEEE/MIT/Amazon Anonymized Network Sensing Graph Challenge, targeting a constant injection rate of $10^5$ requests per second (RPS). Testing was conducted on the MIT SuperCloud using up to 64 compute nodes.

Key Results

  • Throughput: A single DBOS instance successfully sustained over $10^5$ RPS, exceeding the requirements of most real-world web deployments.
  • Scalability: The Python-based implementation exhibited linear scalability up to 64 nodes. The PostgreSQL version scaled linearly up to 32 nodes.
  • Latency: Global analysis was completed in approximately 1.9 seconds for the Python workflow and 12.9 seconds for the PostgreSQL workflow.
  • Overhead: The computational cost of maintaining this level of awareness was found to be negligible relative to the total system resources.

Conclusion and Future Work

By embedding sensing into the OS/Database layer, developers can achieve "security as a side effect." The use of hypersparse matrices provides a privacy-preserving mechanism to share attack "shapes" (e.g., DDoS patterns) without exposing sensitive IP addresses.

Future work will focus on addressing single-threaded bottlenecks in worker nodes, further optimizing the PostgreSQL implementation, and integrating formal prefix-preserving anonymization (e.g., CryptoPAN) into the DBOS architecture. Finally, validating the system against heterogeneous, real-world traffic remains a priority for production readiness.

Summary updated

BibTeX
@inproceedings{lockton2025dbos,
  author = {Sophia Lockton and Jeremy Kepner and Michael Stonebraker and Hayden Jananthan and LaToya Anderson and William Arcand and David Bestor and William Bergeron and Alex Bonn and Daniel Burrill and Chansup Byun and Timothy Davis and Vijay Gadepally and Michael Houle and Matthew Hubbell and Michael Jones and Piotr Luszczek and Peter Michaleas and Lauren Milechin and Chasen Milner and Guillermo Morales and Julie Mullen and Michel Pelletier and Alex Poliakov and Andrew Prout and Albert Reuther and Antonio Rosa and Charles Yee and Alex Pentland},
  title = {{DBOS Network Sensing: A Web Services Approach to Collaborative Awareness}},
  booktitle = {HPEC},
  year = {2025},
  doi = {10.1109/hpec67600.2025.11196670},
}

Loading related content...

Figure