Disk manager with multiple NVMe SSDs #14
Description
One of the main insights of the German's paper is that modern software is unable to exploit the full parallelism of modern hardware, specifically NVMe SSDs that use several PCIe lanes to get crazy throughput.
Right now, the buffer pool manager only supports 1 disk and 1 file. The DiskManager should be an interface that completely abstracts away the number of SSDs the hardware actually has. This also means that the DiskManager should be the one to set up an io_uring instance per disk per thread. For example, if there are 10 SSDs and N worker threads, the DiskManager needs to manage 10 io_uring instances per thread, which means 10 * N io_uring instances.
I don't really know what this means for implementation, or if this is even possible. However, it seems like the Germans were able to get it to work, so theoretically we should be able to as well.