The current version of tired allocation has some
problems.

- If the most common failure reason is N+1 redundancy
  (which it usually is for shared storage, as we take
  global N+1 redundancy into account), no canonical
  shrinking is done.

- It sometimes takes too long to compute the tiered
  allocation statistics.

This patch series addresses the first issue by assuming
that N+1 failures are most likely related to memory. It
improves on the second by only searching for side branches
at reasonable sizes.


Klaus Aehlig (3):
  Instead of sorting and reversing, sort by inverse order
  tieredAlloc: in case of FailN1, shrink memory
  tieredAlloc: Shrink for effect

 src/Ganeti/HTools/Cluster.hs | 46 +++++++++++++++++++++++++++++++++-----------
 1 file changed, 35 insertions(+), 11 deletions(-)

-- 
2.5.0.rc2.392.g76e840b

Reply via email to