io_uring: more disk throughput with less CPU
How Riffle moved local shuffle I/O onto submission and completion queues, added zero-copy reads, and validated the result on a 2.2 TB TeraSort run.
Read articleBlog
Design notes, performance investigations, and implementation details from the Riffle project.
Read ahead: cutting shuffle-read stalls with the Linux page cache
How Riffle recognizes upcoming reads, warms the page cache, and extends prefetching from sequential partitions to AQE read plans.
Read article