In search of a faster SQLite

submited by

Style Pass

2024-12-14 16:30:04

SQLite is already fast. But can we make it even faster? Researchers at the University of Helsinki and Cambridge began with this question and published a paper, “Serverless Runtime / Database Co-Design With Asynchronous I/O”. They demonstrate up to a 100x reduction in tail latency. These are my notes on the paper.

Note that this is a workshop paper and a very short read. It also focuses more on serverless and edge computing. However, the learnings here can be applied more broadly.

The io_uring subsystem in the Linux kernel provides an interface for asynchronous I/O. Its name has its roots in the ring buffers shared between user space and kernel space that alleviate the overheads of copying buffers between them [3]. It allows the application to submit an I/O request and concurrently perform other tasks until it receives a notification from the OS on the com- pletion of the I/O operation.

With io_uring, the application first calls the io_uring_setup() system call to set up two memory regions: the submission and completion queues. The applications then submit I/O requests to the submission queue and call the io_uring_enter() system call to tell the OS to start processing the I/O requests. However, unlike the blocking read() and write()calls, io_uring_enter() does not block the thread by returning control to the userspace. The application can now perform other work concurrently and periodically poll the completion queue in userspace for I/O completion.