Doctoral Thesis Proposal - Sanjith Athlur
November 19, 2026 2:00PM—3:30PM
Location:
4305
-
Newell-Simon Hall
Speaker:
SANJITH ATHLUR,
Ph.D. Student, Computer Science Department, Carnegie Mellon University
Large distributed storage systems store exabytes of data at low cost on mechanical hard disk drives (HDDs). As storage demand grows, HDDs are becoming denser and growing in capacity. However, per-drive IOPS and bandwidth have remained roughly constant. As a result, the IOPS and bandwidth available per terabyte are falling sharply. Storage systems are therefore approaching an IO wall, where workload IO demand outstrips disk IO supply and limits their ability to use denser drives. To keep scaling, they must use disk bandwidth more efficiently and reduce workload IO demand.
This thesis argues that exposing inherent flexibility in how data is stored and accessed can improve IO efficiency and reduce IO demand to help scale the IO wall without changing the useful work performed.
First, Okapi exposes flexibility in how data is stored. Conventional cluster file systems couple how file data is striped across disks to how blocks are grouped for erasure coding. Okapi decouples these entities, allowing applications to indepdently optimize data stripes for IO performance and redundancy groups for durability and space savings. Decoupling also enables redundancy scheme changes to occur without rewriting file data. Okapi improves read throughput by up to 80%, reduces seeks by up to 70%, and cuts redundancy-transition IO by up to 70%.
Second, Declarative IO exposes flexibility in how data is read. Current imperative interfaces (e.g., GET/PUT, read/write) do not expose flexibility. Each imperative request is for a specific data unit to be accessed now. Declarative IO introduces new storage interfaces that allow time- and order-flexible workloads to declare what data they want to access and by when, allowing the storage system to coordinate redundant disk accesses across tasks to effectively reduce disk IO. Declarative IO reduces maintenance disk reads by 28–51% in datacenter-scale workloads, enabling adoption of 1.7× larger HDDs.
Finally, we propose scaling Declarative IO to multi-exabyte clusters while preserving its IO benefits, and extending it to batch analytics workloads by co-scheduling compute, and memory resources with IO reuse.
Thesis Committee:
Greg Ganger (Co-chair)
Rashmi Vinayak (Co-chair)
George Amvrosiadis
Jai Menon (Microsoft)
Contact
Matt Stewart