On Oct 16, 2013, at 1:50 PM, Richard W.M. Jones <rjones@xxxxxxxxxx> wrote: > On Wed, Oct 16, 2013 at 01:09:25PM -0600, Chris Murphy wrote: >> Caching mode was none in all prior cases. > > Note that cache=none is almost never useful. It bypasses the host > cache so you don't get the benefit of having lots of free host memory, > plus causes endless pain because of filesystems that don't support > O_DIRECT (eg. tmpfs) or have buggy O_DIRECT support (eg. glibc on top > of older kernels). It took me quite a long time to realize this -- it > was the default in libguestfs until quite recently :-( Curious, it seems no cache should perform almost as well as writeback, and much better than writethrough. But safer than either writeback or unsafe. I think the qcow2 metadata preallocation is what's making a bulk of the difference in what I'm seeing. qcow2 prealloc lazy metadata1.1 + virtio + unsafe caching is on par with linear LV, which is about as simple a layer as there is. This qcow2 config very slightly beat the LV configuration (also using virtio + unsafe caching); this despite it being effectively ext4 on top of XFS. > > Just remember that cache=unsafe is called unsafe for a reason. If > your host crashes, you can scramble the guest filesystems so it's > impossible to recover. For testing you probably don't care … Nope. So is a compromise between performance and some safety to use writeback instead of unsafe? Might be fun to willfully injure a btrfs file system with both and see if it can at least allow reading of what was successfully written, or if it's just trashed. Chris Murphy -- devel mailing list devel@xxxxxxxxxxxxxxxxxxxxxxx https://admin.fedoraproject.org/mailman/listinfo/devel Fedora Code of Conduct: http://fedoraproject.org/code-of-conduct
