https://github.com/ceph/ceph/pull/869
has a bunch of pending changes to CRUSH to support the erasure coding work in firefly. The main item is that the behavior of 'choose indep' has changed significantly. This is strictly speaking a change in behavior, but nobody should be using indep mode in a normal ceph cluster (unless they went manually fiddling with their crush map). The new and improved indep does a breadth-first mapping instead of depth-first, which means few items shifting around when there are failures. It also drops some of the cruft that fell out of the combined code from before. As a bonus, the old method is now firstn-only and I was able to strip out a bunch of crap in the process. Yay! There are a few other things: - The 'osd crush rule create-simple ..' command now takes an optional mode (firstn or indep) so that it can be used for erasure pools. - There is an 'erasure' pg pool type (existing types were 'rep' (default) and 'raid4' (never used or implemented)). - New rule commands: step set_choose_tries N This overrides the tunable total_tries (default is 50) for the current rule only. step set_chooseleaf_tries M This overrides the recursive behavior when using chooseleaf. By default, for indep mode, we try exactly once with the recursive call, as this maintains the same bound on computational complexity. However, increasing this a bit (say, to 5) improves stability of the mapping a bit when there are devices marked out. This lets you set it for *just* the current rule. Note that for the 'firstn' mode, the default (legacy) behavior is to try total_tries in the recursive call, which makes the computational complexity proprotional to total_tries^2 (in the extreme). If the 'descend_once' tunable is set (now the default), then we do one attempt.. if we hit a reject. Unfortunately not in the case of a collision (dup). But, we can't change that without breaking compatibility for existing rules. To "fix" that, we can add a set_chooseleaf_tries 1 command to firstn rules. It's a bit muddled, though. :( - CrushWrapper has a helper to detect if any of these rule commands are in use, and OSDMap sets the required features accordingly. - There is a small fix for OSDMap CACHEPOOL feature detection. Long story short: if any of this new stuff is used (and it will be needed for erasure pools), the new feature bit will be required and old clients won't be able to connect. I think the new behavior is good. My main concern is the weird interplay of the 'descend_once' tunable, which unfortunately wasn't implemented to mean the same as chooseleaf_tries = 1. I'm not sure if it's worth fixing that via _another_ tunable or not; if so, we could (yay) end up where set_chooseleaf_tries actually works for firstn the same way it does for indep, and the tunable just makes it default to 1 (as it does with indep). sage -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
