Io_uring Without Readahead

The benchmark shows that enabling a 32-page application-level readahead in Turgo’s io_uring backend reduces the number of device read requests from 23 k to about 16 k by submitting many SQEs concurrently, allowing the block layer to merge adjacent reads. In the TPC-H Q6 scan, this concurrency raises throughput, while the sqpoll thread consumes 8 s wall time, 3.7 s user time and 8.5 s system time.