On 10/17/2015 10:28 AM, Michael Torrie wrote: > On 10/17/2015 02:08 AM, Dan Egli wrote: >> Okay, I will just ask this part, and then I think I've gotten all the info >> I need, thanks. You mentioned changing the size of your extent on your >> Dual-Tuner MythTV box. What's the default extent size, and given the usage >> I mentioned (10-20% executable & data - which is where I was thinking of >> Data Dedup, 80-90% multi-media), would that be sufficent for my needs, do >> you think? If not, what would be a good extent size? Given the average size >> of some of the files I will be handling, I'm ALMOST tempted to make an >> extent size of 90MB, but that would be space wasting for the binary >> programs and such. > Hopefully Daniel can speak to his XFS settings. I'm curious about that > myself as my reading of the XFS docs suggests that the only influence > you have with extents on XFS is on something called a real-time > partition, which I don't fully understand, but which seems to be a > special case.
My memory must have bitrot; I confused extents with speculative preallocation using allocsize= option. XFS (like many other Linux systems) tries very hard to use sane values up front. Just as premature optimization is the root of all evil, so is premature XFS tuning. If you need to tune it further, tune later. Frankly, you are more likely to hurt performance by tuning it, especially if you pvmove it to something different and all the assumptions change. I only tuned my video pools because I was using slow SATA drives that were not keeping up with journaling, metadata, and two encoder outputs, buffering live video, playing recorded video, and frequent disk syncs. The 64MB buffer/file prealloc really helped keep the fragmentation and head seeks down, and probably helped with interrupts as well, improving my overall throughput. Since then my assumptions have changed; I no longer have analog cable, I don't want to pay through the nose for digital cable, and it seems there is nothing on TV worth watching anyway. The box now sits lonely in the corner waiting for me to do something different with it. Last I heard, the real-time partition is not implemented on Linux; it was an Irix only thing, and was only supported on special hardware. Metadata performance has been slow on XFS in the past (deleting/modifying attributes of millions of small files), but I've heard this has changed in newer kernels. I haven't tested it though. Grazie, ;-Daniel Fussell /* PLUG: http://plug.org, #utah on irc.freenode.net Unsubscribe: http://plug.org/mailman/options/plug Don't fear the penguin. */
