{-
-Copyright (C) 2009 Google Inc.
+Copyright (C) 2009, 2010 Google Inc.
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
-- * Allocation functions
, iterateAlloc
, tieredAlloc
+ , instanceGroup
+ , findSplitInstances
+ , splitCluster
) where
import Data.List
, (1, "mem_load_cv")
, (1, "disk_load_cv")
, (1, "net_load_cv")
- , (1, "pri_tags_score")
+ , (2, "pri_tags_score")
]
detailedCVWeights :: [Double]
mem_cv = varianceCoeff mem_l
-- metric: disk covariance
dsk_cv = varianceCoeff dsk_l
- n1_l = length $ filter Node.failN1 nodes
- -- metric: count of failN1 nodes
- n1_score = fromIntegral n1_l::Double
+ -- metric: count of instances living on N1 failing nodes
+ n1_score = fromIntegral . sum . map (\n -> length (Node.sList n) +
+ length (Node.pList n)) .
+ filter Node.failN1 $ nodes :: Double
res_l = map Node.pRem nodes
-- metric: reserved memory covariance
res_cv = varianceCoeff res_l
allocateOnSingle nl inst p =
let new_pdx = Node.idx p
new_inst = Instance.setBoth inst new_pdx Node.noSecondary
- new_nl = Node.addPri p inst >>= \new_p ->
- return (Container.add new_pdx new_p nl, new_inst, [new_p])
- in new_nl
+ in Node.addPri p inst >>= \new_p -> do
+ let new_nl = Container.add new_pdx new_p nl
+ new_score = compCV nl
+ return (new_nl, new_inst, [new_p], new_score)
-- | Tries to allocate an instance on a given pair of nodes.
allocateOnPair :: Node.List -> Instance.Instance -> Node.Node -> Node.Node
allocateOnPair nl inst tgt_p tgt_s =
let new_pdx = Node.idx tgt_p
new_sdx = Node.idx tgt_s
- new_nl = do -- Maybe monad
- new_p <- Node.addPri tgt_p inst
- new_s <- Node.addSec tgt_s inst new_pdx
- let new_inst = Instance.setBoth inst new_pdx new_sdx
- return (Container.addTwo new_pdx new_p new_sdx new_s nl, new_inst,
- [new_p, new_s])
- in new_nl
+ in do
+ new_p <- Node.addPri tgt_p inst
+ new_s <- Node.addSec tgt_s inst new_pdx
+ let new_inst = Instance.setBoth inst new_pdx new_sdx
+ new_nl = Container.addTwo new_pdx new_p new_sdx new_s nl
+ return (new_nl, new_inst, [new_p, new_s], compCV new_nl)
-- | Tries to perform an instance move and returns the best table
-- between the original one and the new one.
tryBalance :: Table -- ^ The starting table
-> Bool -- ^ Allow disk moves
-> Bool -- ^ Only evacuate moves
+ -> Score -- ^ Min gain threshold
+ -> Score -- ^ Min gain
-> Maybe Table -- ^ The resulting table and commands
-tryBalance ini_tbl disk_moves evac_mode =
+tryBalance ini_tbl disk_moves evac_mode mg_limit min_gain =
let Table ini_nl ini_il ini_cv _ = ini_tbl
all_inst = Container.elems ini_il
all_inst' = if evac_mode
fin_tbl = checkMove node_idx disk_moves ini_tbl reloc_inst
(Table _ _ fin_cv _) = fin_tbl
in
- if fin_cv < ini_cv
+ if fin_cv < ini_cv && (ini_cv > mg_limit || ini_cv - fin_cv >= min_gain)
then Just fin_tbl -- this round made success, return the new table
else Nothing
concatAllocs :: AllocSolution -> OpResult Node.AllocElement -> AllocSolution
concatAllocs (flst, cntok, sols) (OpFail reason) = (reason:flst, cntok, sols)
-concatAllocs (flst, cntok, osols) (OpGood ns@(nl, _, _)) =
- let nscore = compCV nl
- -- Choose the old or new solution, based on the cluster score
+concatAllocs (flst, cntok, osols) (OpGood ns@(_, _, _, nscore)) =
+ let -- Choose the old or new solution, based on the cluster score
nsols = case osols of
[] -> [(nscore, ns)]
(oscore, _):[] ->
let em = do
(mnl, i, _, _) <-
applyMove nl inst (ReplaceSecondary x)
- return (mnl, i, [Container.find x mnl])
+ return (mnl, i, [Container.find x mnl],
+ compCV mnl)
in concatAllocs cstate em
) ([], 0, []) valid_idxes
in return sols1
-- FIXME: hardcoded one node here
(fm, cs, aes) <- tryReloc nl' il idx 1 ex_ndx
case aes of
- csol@(_, (nl'', _, _)):_ ->
+ csol@(_, (nl'', _, _, _)):_ ->
return (nl'', (fm, cs, csol:rsols))
_ -> fail $ "Can't evacuate instance " ++
- show idx
+ Instance.name (Container.find idx il)
) (nl, ([], 0, [])) all_insts
return sol
-> Instance.Instance
-> Int
-> [Instance.Instance]
- -> Result (FailStats, Node.List, [Instance.Instance])
+ -> Result (FailStats, Node.List, Instance.List,
+ [Instance.Instance])
iterateAlloc nl il newinst nreq ixes =
let depth = length ixes
newname = printf "new-%d" depth::String
Bad s -> Bad s
Ok (errs, _, sols3) ->
case sols3 of
- [] -> Ok (collapseFailures errs, nl, ixes)
- (_, (xnl, xi, _)):[] ->
- iterateAlloc xnl il newinst nreq $! (xi:ixes)
+ [] -> Ok (collapseFailures errs, nl, il, ixes)
+ (_, (xnl, xi, _, _)):[] ->
+ iterateAlloc xnl (Container.add newidx xi il)
+ newinst nreq $! (xi:ixes)
_ -> Bad "Internal error: multiple solutions for single\
\ allocation"
-> Instance.Instance
-> Int
-> [Instance.Instance]
- -> Result (FailStats, Node.List, [Instance.Instance])
+ -> Result (FailStats, Node.List, Instance.List,
+ [Instance.Instance])
tieredAlloc nl il newinst nreq ixes =
case iterateAlloc nl il newinst nreq ixes of
Bad s -> Bad s
- Ok (errs, nl', ixes') ->
+ Ok (errs, nl', il', ixes') ->
case Instance.shrinkByType newinst . fst . last $
sortBy (comparing snd) errs of
- Bad _ -> Ok (errs, nl', ixes')
+ Bad _ -> Ok (errs, nl', il', ixes')
Ok newinst' ->
- tieredAlloc nl' il newinst' nreq ixes'
+ tieredAlloc nl' il' newinst' nreq ixes'
-- * Formatting functions
ReplaceSecondary ns -> [ opR ns ]
ReplaceAndFailover np -> [ opR np, opF ]
FailoverAndReplace ns -> [ opF, opR ns ]
+
+-- | Computes the group of an instance
+instanceGroup :: Node.List -> Instance.Instance -> Result GroupID
+instanceGroup nl i =
+ let sidx = Instance.sNode i
+ pnode = Container.find (Instance.pNode i) nl
+ snode = if sidx == Node.noSecondary
+ then pnode
+ else Container.find sidx nl
+ puuid = Node.group pnode
+ suuid = Node.group snode
+ in if puuid /= suuid
+ then fail ("Instance placed accross two node groups, primary " ++ puuid ++
+ ", secondary " ++ suuid)
+ else return puuid
+
+-- | Compute the list of badly allocated instances (split across node
+-- groups)
+findSplitInstances :: Node.List -> Instance.List -> [Instance.Instance]
+findSplitInstances nl il =
+ filter (not . isOk . instanceGroup nl) (Container.elems il)
+
+-- | Splits a cluster into the component node groups
+splitCluster :: Node.List -> Instance.List ->
+ [(GroupID, (Node.List, Instance.List))]
+splitCluster nl il =
+ let ngroups = Node.computeGroups (Container.elems nl)
+ in map (\(guuid, nodes) ->
+ let nidxs = map Node.idx nodes
+ nodes' = zip nidxs nodes
+ instances = Container.filter ((`elem` nidxs) . Instance.pNode) il
+ in (guuid, (Container.fromAssocList nodes', instances))) ngroups
projects / ganeti-local / blobdiff