CAREER: In-Kernel Execution of Storage Functions

Emerging hardware storage devices can achieve millions of operations per second. They can also access data in as little as 2-3 microseconds, representing an order of magnitude improvement over legacy flash memory and hard disk technologies. However, at such low latencies, the overhead of the operating and file system becomes a significant bottleneck to end-to-end storage performance. Storage requests may slow down by 2x or more. The project designs XRP (eXpress Resubmission Path), a Linux-supported framework that allows applications to write storage functions that can be executed by the operating system. XRP allows applications to bypass almost entirely the overhead of the operating and file system.

The main research challenge is that by bypassing the operating system's storage layers, XRP lacks important context about the user's request, such as who owns the block it is accessing, and how to traverse the application's data structure safely and concurrently. The project organizes four primary research thrusts: (a) adding support for XRP in the Linux kernel and file system, (b) supporting a variety of storage operations and integrating with real-world storage systems, (c) enforcing quality of service, isolation and fairness of in-kernel storage functions, and (d) implementing in-kernel handling of over-the-network storage requests.

Fast storage devices will fundamentally change the way software systems are built, and will require a redesign of the entire storage stack underpinning these systems. By eliminating the software overhead when accessing such devices, while maintaining operating system support, XRP will enable systems to achieve both high throughput and low and predictable tail latency, while leaving intact the existing isolation, security and safety guarantees of the Linux kernel. The immediate benefits will include high-performing datacenter data systems at a reduced cost, which will in turn translate to cost-effective cloud services for end users.

Software artifacts will be released under open-source licenses to enable other researchers, and work with hardware and cloud providers to transition our research into practice. The project repository will be available via the principal investigator's website: https://www.asafcidon.com

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.