On 05/16/2014 04:41 PM, Tomasz Chmielewski wrote: > On Fri, 16 May 2014 14:06:24 -0400 > Calvin Walton wrote: > >> No comment on the performance issue, other than to say that I've seen >> similar on RAID-10 before, I think. >> >>> Also, what happens when the system crashes, and one drive has >>> several hundred megabytes data more than the other one? >> >> This shouldn't be an issue as long as you occasionally run a scrub or >> balance. The scrub should find it and fix the missing data, and a >> balance would just rewrite it as proper RAID-1 as a matter of course. > > It's similar (writes to just one drive, while the other is idle) when > removing (many) snapshots. > > Not sure if that's optimal behaviour. > I think, after having looked at some of the code, that I know what is causing this (although my interpretation of the code may be completely off target). As far as I can make out, BTRFS only dispatches writes to one device at a time, and the write() system call only returns when the data is on both devices. While dispatching to one device at a time is optimal when both 'devices' are partitions on the same underlying disk (and also if your optimization metric is the simplicity of the underlying code), it degrades very fast to the worst case when using multiple devices. The underlying cause however, which the one device at a time logic in BTRFS just makes much worse, is that the buffer for the write() call is kept in memory until the write completes, and counts against the per-process write-caching limit, and when the process fills up it's write-cache, the next call it makes that would write to the disk hangs until the write cache is less full. The two options that I've found that work around this are: 1. Run 'sync' whenever the program stalls, or 2. Disable write-caching by adding the following to /etc/sysctl.conf vm.dirty_bytes = 0 vm.dirty_background_bytes = 0 Option 1 is kind of tedious, but doesn't hurt performance all that much, Option 2 will lower throughput, but will cause most of the stalls to disappear. Ideally, BTRFS should dispatch the first write for a block in a round-robin fashion among available devices. This won't fix the underlying issue, but it will make it less of an issue for BTRFS.