Thread by @zhong_yuhong

Very excited that XRP wins Jay Lepreau Best Paper Award 🏆 at #osdi22! XRP (open-sourced) is the first system that adopts #Linux #eBPF 🐝 to reduce kernel software overhead for storage. (1/9)

With new storage technology like 3D XPoint, storage devices are getting much faster ⚡️. As a result, kernel software is becoming the bottleneck for storage! See the breakdown of the average latency of random 512B reads. (2/9)

One extreme way to address the kernel software overhead is to bypass the kernel. However, kernel bypass comes with several issues 💣, such as no fine-grained access control and poor CPU utilization. (3/9)

In order to reduce the kernel software overhead without bypassing the kernel, we build XRP, a framework that allows applications to offload custom eBPF functions into the NVMe driver to submit/process storage I/Os. (4/9)

Think about a B+ tree index. Without XRP, an index lookup consists of issuing multiple read requests to walk down the B+ tree from user space, which traverses the kernel software stack multiple times. (5/9)

With XRP, the eBPF function can parse B+ tree nodes and submit more requests right at the NVMe driver, which means we only need to traverse the kernel software stack once during the entire index lookup! (6/9)

We demonstrate that XRP can accelerate MongoDB's WiredTiger and can support many types of operations (e.g., point queries, range queries, aggregations). (7/9)

We are very excited about XRP and are integrating XRP with other key-value stores such as RocksDB. Try it out and see if it can make your database faster⚡️: github.com/xrp-project/XRP (8/9)

XRP is the result of an awesome team effort. A big shout out to my advisors and collaborators: Haoyu Li, Yu Jian Wu, @yanniszark, Jeffrey Tao, @EvanMesterhazy, Michael Makris, @junfengy, Amy Tai, @rstutsman, and @asafcidon! (9/9)