On 2017-07-21 19:21, Hans van Kranenburg wrote: > On 07/21/2017 05:50 PM, Austin S. Hemmelgarn wrote: >> On 2017-07-21 07:47, Hans van Kranenburg wrote: >>> [...] >>> >>> Signed-off-by: Hans van Kranenburg <hans.van.kranenb...@mendix.com> >> Behaves as advertised, and I'm not seeing any issues in my testing (and >> I can confirm from experience that this logic slows things down, I've >> been forcing '-o nossd' on my systems for a while now for the >> performance improvement), so you can add: >> Tested-by: Austin S. Hemmelgarn <ahferro...@gmail.com> > > Thanks for testing! > > Would you mind sharing a bit about what kind of test this is, and how > you measure the difference? > > Does it mean that you're throwing a lot of data at the disk, and that > you can measure that 'cluster' mode of figuring out where to put the > writes has to spend more effort than the 'tetris' mode, and that it > limits the write speed? > > The only relevant difference I was aware of until now is just the > placement of the writes. >
For the generic testing, I've got a (mostly) automated testing system setup on my home server that builds a kernel with the patches applied and then runs xfstests and a handful of other tests I've got myself (I really need to get around to pushing those upstream...). At the moment, it tests using 32 and 64-bit x86, ARMv7a, ARMv6stj-z, ARMv8 (whatever QEMU's default revision is), PPC64 (big and little endian), and 32-bit MIPS (new ABI only, also big and little endian). For the specific case here, back when I set up my systems to use 'nossd' 3 months ago, I specifically set up two of them that are otherwise identical to test if there was any difference. The systems in question are a pair of Intel NUC's with Pentium N3700 CPU's (4 cores, no HT, 1.6GHz normal, 2.4GHz boost), 4G of DDR3L-1600 RAM, and a 250GB Crucial MX200 SSD. After 3 months, the one which has run with 'nossd' set in the mount options is getting roughly 1-3% better write performance, equivalent read performance (both measured with fio and iozone), and completes discard operations almost twice as fast (based on timing `fstrim` with the same amount of space free on both filesystems). -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html