Support for fuse max read option

[llssyy1991]

The fuse inside the mountpoint-s3 is using the default max_read which is 128K per call. For our scenario, we are using the mount-s3 to stream some large data set and tarballs, usually in tens of Gbs, we found there are certain of the time is consumed by the large amound of fuse calls. In order to better improve the performance, we would like to have some option from mount-s3 to modify the max_read internally.

[vladem]

Thanks for opening the issue! The maximum size of a single READ request that the kernel sends to the FUSE daemon is configured in the kernel code and equals to 32 pages (128KiB assuming 4KiB page size) for non direct IO and 256 pages (1024KiB) when opened with O_DIRECT:

• https://github.com/torvalds/linux/blob/master/fs/fuse/fuse_i.h#L36
• https://github.com/torvalds/linux/blob/86731a2a651e58953fc949573895f2fa6d456841/fs/fuse/inode.c#L39

You can use the O_DIRECT flag when opening files, which will increase the read size from 128KiB to 1024KiB. Additionally, while the max_readahead is set to 128KiB (kernel default value for non direct IO), if the application calls posix_fadvise with POSIX_FADV_SEQUENTIAL, the kernel will double the read size.

We'll continue to explore ways to improve performance for large file operations. Please share any additional performance metrics or specific use cases that could help us better understand your requirements.