On Tue, 2006-05-16 at 12:37 -0400, Greg Freemyer wrote:
If you read it, you will see that the XFS team basically say the performance of XFS would be greatly degraded if they eliminated the window of opportunity. They describe the window as possibly being minutes long between a Meta-Data update that increases the filesize and the actual data being flushed to disk.
I think people forget that all meta-data journaling filesystems are a set of hacks. They do *NOT* improve the integrity of your filesystem. They are *NOT* "true atomic commits." The "main reason" for a meta-data journaling filesystem is to reduce or eliminate the filesystem integrity check. The filesystem integrity check merely puts the filesystem in a "consistent" state, and does *NOT* necessarily mean you didn't have data loss. If you are "anal" on integrity, but want to reduce recovery time to make a filesystem inconsistent, the "most reliable" approach is as follows ... 1. Filesystem with full data journaling 2. Non-volatile RAM (NVRAM) for full data journal 3. Firmware "supervisor" for NVRAM/disk buffer flush/recovery 4. An integrated volume management and filesystem #1-4 is essentially what NetApp's Data OnTap OS and WAFL integrated volume management and filesystem does. It not only does full data journaling, but uses a NVRAM board and a firmware "supervisor" to ensure all NVRAM/disk buffer is flushed/recovered in the case of a system crash, loss of power, etc... Now prior to its leaving the hardware business, VA Linux / Research _did_ sell Linux 2.2 solutions with early, full data journaling-only Ext3. They also included a NVRAM board for the data journal. This was #1-2. It came as close as you could get to a "most reliable" setup. It essentially _eliminated_ non-atomic commits as well as put the data to the NVRAM quickly. For those of us running the NFS v3 SGI/Trond/Higgen patches on late 2.2 kernels at the time, having this solution (Ext3 full data + NVRAM) meant we could run a _true_ NFS v3 sync operation that was just as fast as Linux's non-standards compliant NFS v2 async (which is not a valid IETF NFS v2 mode). It was a major blessing. Today, using a full data journal and NVRAM is largely overkill. But if you want the ultimate in "reliability/integrity," it's still not a bad idea. Unfortunately, Ext3 is now meta-data and the full data journaling code virtually hasn't been worked on since kernel 2.2 (i.e., I don't know if I trust it). -- Bryan J. Smith Professional, technical annoyance mailto:b.j.smith@ieee.org http://thebs413.blogspot.com ----------------------------------------------------------- Americans don't get upset because citizens in some foreign nations can burn the American flag -- Americans get upset because citizens in those same nations can't burn their own