SDI / Parallel Data Laboratory Talks - Virkramraj Sitpal, Shubham Kumar, Rajarshi Chowdhury
June 18, 2026 12:00PM—2:00PM
Location:
Virtual Presentations - ET
-
Remote Access Zoom
Speaker:
VIKRAMRAJ SITPAL, SHUBHAM KUMAR, & RAJARSHI CHOWDHURY,
Vikramraj, Senior Development Manager; Subham, Senior Member of Technical Staff; Rajarshi, Consulting Member of Technical Staff Oracle
TALK 1
Vikramraj Sitpal & Shubham Kumar, Oracle
— Why We Created Yet Another Memory Framework: Understanding MGA’s Role in Next-Gen Database Systems
Modern database engines already provide multiple memory areas, but under production constraints they often force a difficult choice: use globally shared memory (such as the SGA), with coarse visibility and lifecycle semantics, or use private process memory (such as the PGA), with stronger isolation but more copying, fragmentation, and indirect coordination when state must be shared selectively. This talk presents Managed Global Area (MGA) in Oracle AI Database as a scoped shared-memory abstraction that fills this missing middle. MGA lets database components define the allocation source, process membership, synchronization policy, lifecycle, and recovery behavior for memory shared across selected processes, without imposing system-wide visibility. We will describe how MGA is built around namespaces, segments, address-space reservation, dynamic attach/detach, fixed variables, recovery records, and observability, and show how these mechanisms make shared memory usable in a production database setting. Through analytical and AI workloads that stress shared-memory execution, our work demonstrates that dynamically scoped shared memory can improve both efficiency and predictability, reducing join-intensive TPC-H latency by up to 35%, model-sharing memory footprint by up to 90%, and large-model inference latency by up to 37%.
Vikramraj Sitpal is a Senior Development Manager at Oracle, focusing on database kernel development within the Database organization. He has had the opportunity to work on various code infrastructure modules, including memory, synchronization, resource management, code sandboxing and I/O. With experience across operating systems, storage systems, networks, distributed systems, and database technologies, he enjoys collaborating with others to solve challenging technical problems and contribute to Oracle’s ongoing innovation. He holds a Master’s degree in Computer Systems from Carnegie Mellon University (MSIN ’21)
Shubham Kumar is a Senior Member of Technical Staff in the Virtual Operating Systems team at Oracle. His work sits at the intersection of database internals, operating systems, and hardware, with a focus on memory management, memory isolation, and observability in large-scale database systems. He is especially interested in making production memory behavior more predictable, efficient, secure, and easier to diagnose. He has a Bachelor's degree from Indian Institute of Information Technology, Ranchi and a Master's degree from State University of New York, Binghamton.
TALK 2:
Rajarshi Chowdhury, Oracle
— Rethinking Database I/O: io_uring, Atomic Writes, and the Road Ahead
Modern NVMe devices sustain millions of IOPS, but the traditional Linux I/O stack burns CPU cycles a database would rather spend on query processing. This talk describes the integration of io_uring into Oracle Database 26ai's storage layer at production scale, and why selective adoption, rather than wholesale replacement, is the right strategy for a multi-process, multi-threaded application. Microbenchmark evaluations lead us to a hybrid architecture built on per-process ring contexts, transparent libaio fallback, and a shared buffer registration mechanism that Oracle initiated and is now in mainline Linux — delivering substantially higher write throughput and lower CPU per I/O on production workloads. We then turn to data integrity: the same io_uring path now can carry atomic (untorn) writes, bringing storage-level torn-write protection to the database. We close with the road ahead: NVMe passthrough, and the longer arc toward programmable, device-direct I/O.
Rajarshi Chowdhury is a Consulting Member of Technical Staff at Oracle, where he works in the Virtual Operating System (VOS) group, the layer where the database meets the operating system. Over seventeen years on the Oracle Database kernel, his focus has been the I/O and storage path, resource management, RDMA fabrics, and the Linux kernel, with particular attention to how thousands of consolidated tenants perform predictably under load. His recent work reaches further up the stack into query processing, GPU offload for data-intensive and AI workloads, and the growing overlap of AI and databases. He holds an M.Tech in Computer Science & Engineering from IIT Roorkee.
Zoom Participation. See announcement for details or to participate in the seminar s
For More Information:
karen@ece.cmu.edu