PublishedI/O for LLM Inference: A Survey of Storage and Memory Bottlenecks
@article{chowdhury2026ioforllm, author = {Chowdhury, Rajarshi}, title = {{I/O} for {LLM} Inference: A Survey of Storage and Memory Bottlenecks}, journal = {Artificial Intelligence Review}, volume = {59}, number = {11}, pages = {231}, year = {2026}, publisher = {Springer}, doi = {10.1007/s10462-026-11651-1} } R. Chowdhury, "I/O for LLM Inference: A Survey of Storage and Memory Bottlenecks," Artificial Intelligence Review, vol. 59, no. 11, Art. no. 231, 2026, doi: 10.1007/s10462-026-11651-1.Database Kernel · AI · Systems · Storage & I/O
Rajarshi Chowdhury
Database Kernel Architect — Consulting Member of Technical Staff, Oracle
Redwood City, California
About
I work on the core of the database kernel — operating-system internals, processors and memory, and how data moves between storage, network, and CPU. My work also reaches further up the stack: query processing, GPU offload for data-intensive and AI workloads, and the growing overlap of AI and databases. I like building the kind of low-level systems that quietly hold everything else up, and I write and speak about it at systems, storage, database, and security venues.
Focus Areas
Experience
Consulting Member of Technical Staff
Feb 2017 — Present
Oracle
Principal Member of Technical Staff
Jan 2010 — Feb 2017
Oracle (via Sun Microsystems acquisition)
Software Engineer
Aug 2008 — Jan 2010
Sun Microsystems
Education
M.Tech, Computer Science & Engineering
2008
Indian Institute of Technology (IIT) Roorkee
B.Tech, Computer Science & Engineering
2005
Govt. College of Engineering & Textile Technology
Selected Publications · 2026
PublishedA Secure, Manifest-Based Framework for Delegated Privilege Promotion
A security architecture for automated patching in air-gapped environments.
@inproceedings{chowdhury2026manifest, author = {Chowdhury, Rajarshi and Shah, Akshay}, title = {A Secure, Manifest-Based Framework for Delegated Privilege Promotion}, booktitle = {2026 56th Annual IEEE/IFIP International Conference on Dependable Systems and Networks -- Supplemental Volume (DSN-S)}, pages = {8--14}, year = {2026}, month = jun, address = {Charlotte, NC, USA}, publisher = {IEEE}, eprint = {2605.28991}, archivePrefix = {arXiv} } R. Chowdhury and A. Shah, "A Secure, Manifest-Based Framework for Delegated Privilege Promotion," in Proc. 56th Annu. IEEE/IFIP Int. Conf. Dependable Systems and Networks – Supplemental Volume (DSN-S), Charlotte, NC, USA, 2026, pp. 8–14.PublishedIORM: Hierarchical I/O Governance for Thousands of Consolidated Databases on Oracle Exadata
Control-plane / data-plane architecture for enforcing strict SLAs in multi-tenant environments.
@article{chowdhury2026iorm, author = {Chowdhury, Rajarshi and Shah, Akshay and Alrmaih, Zakaria and Guo, Chenhao and Singh, Anubhav and Lee, Sue}, title = {{IORM}: Hierarchical {I/O} Governance for Thousands of Consolidated Databases on {Oracle Exadata}}, journal = {Proceedings of the VLDB Endowment}, volume = {19}, number = {12}, pages = {4440--4452}, year = {2026}, doi = {10.14778/3827998.3828044} } R. Chowdhury, A. Shah, Z. Alrmaih, C. Guo, A. Singh, and S. Lee, "IORM: Hierarchical I/O Governance for Thousands of Consolidated Databases on Oracle Exadata," Proc. VLDB Endow., vol. 19, no. 12, pp. 4440–4452, 2026, doi: 10.14778/3827998.3828044.AcceptedHardware-Transparent I/O Governance in Disaggregated Heterogeneous Storage
The I/O Resource Manager used in Oracle Exadata / Exascale to resolve "noisy-neighbor" conflicts at hyperscale.
@inproceedings{chowdhury2026hwtransparent, author = {Chowdhury, Rajarshi and Shah, Akshay and Lee, Sue K.}, title = {Hardware-Transparent {I/O} Governance in Disaggregated Heterogeneous Storage}, booktitle = {Proceedings of the 14th IEEE International Conference on Cloud Engineering (IC2E)}, year = {2026}, note = {To appear}, eprint = {2607.16578}, archivePrefix = {arXiv} } R. Chowdhury, A. Shah, and S. K. Lee, "Hardware-Transparent I/O Governance in Disaggregated Heterogeneous Storage," in Proc. 14th IEEE Int. Conf. Cloud Engineering (IC2E), 2026, to appear. arXiv:2607.16578.Earlier Papers
Paperuringscope: Portable, Low-Overhead Observability for io_uring
@misc{chowdhury2026uringscope, author = {Chowdhury, Rajarshi}, title = {uringscope: Portable, Low-Overhead Observability for io\_uring}, year = {2026}, eprint = {2606.15137}, archivePrefix = {arXiv}, primaryClass = {cs.OS}, doi = {10.48550/arXiv.2606.15137} } R. Chowdhury, "uringscope: Portable, Low-Overhead Observability for io_uring," arXiv:2606.15137 [cs.OS], 2026.PaperDesign and Implementation of Parallel Pattern Matching Algorithms for Multi-Core Systems
All citations
Upcoming Conferences & Talks
TalkIORM: Hierarchical I/O Governance for Thousands of Consolidated Databases on Oracle Exadata
TalkHardware-Transparent I/O Governance in Disaggregated Heterogeneous Storage
Earlier Talks
TalkInvited talk — Parallel Data Lab (PDL) Speaker Series
TalkA Secure, Manifest-Based Framework for Delegated Privilege Promotion
TalkIntegrating SPDK with Oracle RDBMS
TalkTop 10 Tips and Tricks for Getting the Best Database Performance from Oracle Solaris Systems
Patents
Hybrid I/O Management Architecture with Tiered Dispatch
Secure Software Component Activation through Signature-Governed Attribute Change
Hyper-Converged Storage Architecture with Zero-Copy NVMe-over-Fabrics for High-Performance Clustered Systems
Open Source
Librarypolicybook
Runnable decision policies for cache eviction, rate limiting, retries, and LLM KV-cache eviction, each with a one-page explainer, shared test vectors, and reproducible benchmarks. Copy a policy into your repo via the CLI (npx policybook add), or install the whole set with pip; C ships as a single header. MIT licensed.
Tooluringscope
Kernel and userspace tracing for io_uring: a single-binary tracer and analyzer that attaches to any process and reconstructs what each request did, from per-opcode latency and hidden async-worker punts to batching efficiency and stalls. Each report ends with findings that name the problem, show the evidence, and suggest a fix; a --check mode catches buffer races and use-after-free hazards, with JSON reports, CI gates, live mode, OpenMetrics, and Perfetto timelines alongside. Built on CO-RE eBPF, so it survives the kernel's io_uring tracepoint churn instead of pinning to one release.
Peer Review & Service
- 2027 Industry Track Co-Chair — DSN 2027
- 2027 Senior Reviewer — SIGCSE TS 2027
- 2026 Industry Session Chair — DSN 2026
- 2026 Reviewer — SIGCSE Virtual
- 2026 Reviewer — ICANN