[Bug 230260] [FUSE] [PERFORMANCE]: Performance issue (I/O block size)

[bugzilla-noreply at freebsd.org]

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=230260

            Bug ID: 230260
           Summary: [FUSE] [PERFORMANCE]: Performance issue (I/O block
                    size)
           Product: Base System
           Version: 11.1-RELEASE
          Hardware: Any
               URL: https://robo.moosefs.com/support/fuse_helloworld.tgz
                OS: Any
            Status: New
          Severity: Affects Some People
          Priority: ---
         Component: kern
          Assignee: bugs at FreeBSD.org
          Reporter: freebsd at moosefs.com

This is one of three issues we detected in FreeBSD FUSE while developing our
distributed file system. All four issues can be replicated using this simple
test script:
https://robo.moosefs.com/support/fuse_helloworld.tgz

Performance issue in FUSE: if a program uses FUSE without the "direct" option,
any I/O is always performed in 4k blocks. Maximum I/O speed we managed to get
was 600MB/s (no physical I/O, just sending zeros from a RAM buffer).

With "direct" it's fast, 5GB/s, but "direct" is not the best solution: no
cache, read operation has no limit on block size and if one uses extremely big
block size, the read speed drastically drops again (we performed dd with bs=1G
and the speed was only 40MB/s). Generally, "direct" is geared toward
stream-like data (character devices) and should not be used for disk-like I/O.

Other FUSE implementations (Linux, MacOS) use 64k block.

Best regards,
Peter / MooseFS Team

-- 
You are receiving this mail because:
You are the assignee for the bug.