Project Overview
Modern cloud infrastructures widely employ storage disaggregation to improve scalability, resource utilization, and cost efficiency. However, today’s disaggregated storage systems continue to follow a “smart-sender, dumb-receiver” design philosophy, where each remote NVMe SSD is exposed as a locally attached block device managed entirely by the client-side storage stack. While this abstraction enables seamless adoption without modifying existing applications, it leaves the storage infrastructure largely passive. As a result, storage systems cannot adapt to dynamic workload interference, coordinate resources across tenants, or exploit the growing programmability of modern storage and networking hardware, leading to unpredictable performance and inefficient resource utilization.
The Programmable Disaggregated Storage (PDS) project builds a new type of disaggregated storage by making the entire storage access path programmable. PDS leverages the recent and upcoming hardware innovations across the networking and storage infrastructures (such as open, computational, and virtualizable SSDs, programmable adapters at the endpoints, and programmable switching fabric in the middle). It co-designs storage software with these hardware substrates to provide adaptive storage abstractions, end-to-end observability, distributed resource management, and programmable I/O processing. The goal is to build storage systems that dynamically adapt to application demands and infrastructure conditions while delivering higher performance, predictable latency, efficient multi-tenancy, and better resource utilization.
Challenges
Emerging storage and networking hardware change how disaggregated storage systems should be built. However, existing software continues to treat NVMe SSDs, storage adapters, and SAN switches as isolated components, leaving their capabilities largely underutilized. Building PDS requires unified abstractions, end-to-end observability, coordinated resource management across heterogeneous devices, and integrated control over the entire storage access path. This incurs several challenges:
- Inflexible storage abstraction, exposing passive block devices instead of elastic storage services.
- Restricted programmability, lacking unified interfaces for coordinating heterogeneous storage and networking devices.
- Limited observability, providing little end-to-end visibility into storage, network, and protocol execution.
- Heterogeneous resource management, requiring coordinated allocation and scheduling of compute, memory, network, and storage resources across diverse programmable devices.
- Fragmented end-to-end control, optimizing individual storage and networking components independently rather than orchestrating the entire storage access path.
Our Approach
Our insight is that the disaggregated storage I/O path can be architected as an end-to-end and reconfigurable active substrate rather than a passive data transport pipeline. We ask: How can PDS leverage emerging hardware innovations to deliver high-performance I/O with predictable latency, efficient multi-tenancy, and cost-effective resource utilization? In this project, we have built systems for storage virtualization, adaptive I/O scheduling, end-to-end orchestration, elastic block storage, NVMe-over-Fabrics observability, and SAN networking fabric, and retrofitted their benefits for distributed file systems, key-value stores and databases.
Here are the systems we have built.
- SANarch compares the switched and switchless architecture for disaggregated SAN (NSDI’26).
- TapDB understands and optimizes database pushdown over storage disaggregation (ASPLOS’26).
- GluonFS ranked 26th in the 10-node category of IO500 and 33rd overall.
- ntprof is a profiling utility for characterizing and analyzing the NVMe-over-TCP protocol (NSDI’25).
- Flint builds an elastic block storage over EBOFs using shadow views (NSDI’25).
- LEED develops a low-power and fast key-value store over SmartNIC JBOFs (SIGCOMM’23).
- eZNS provides an elastic and predictable zoned interface for small-zone ZNS SSDs (OSDI’23).
- Dremel configures the RocksDB adaptively over disaggregated storage (SIGMETRICS’22).
- Gimbal enables efficient multi-tenancy on NVMe-oF targets for SmartNIC JBOFs (SIGCOMM’21).
- fRSM proposes fine-grained replicated state machines for metadata indexes of cluster storage (NSDI’20).