seaweedfs/weed/shell/command_ec_balance.go

540 lines
18 KiB
Go
Raw Normal View History

2019-05-30 08:38:59 +00:00
package shell
import (
"flag"
"fmt"
"io"
"github.com/seaweedfs/seaweedfs/weed/pb"
"github.com/seaweedfs/seaweedfs/weed/storage/erasure_coding"
"github.com/seaweedfs/seaweedfs/weed/storage/needle"
"github.com/seaweedfs/seaweedfs/weed/storage/types"
"golang.org/x/exp/slices"
2019-05-30 08:38:59 +00:00
)
func init() {
Commands = append(Commands, &commandEcBalance{})
2019-05-30 08:38:59 +00:00
}
type commandEcBalance struct {
}
func (c *commandEcBalance) Name() string {
return "ec.balance"
}
func (c *commandEcBalance) Help() string {
2019-06-11 04:32:56 +00:00
return `balance all ec shards among all racks and volume servers
2019-05-30 08:38:59 +00:00
2019-06-05 07:11:37 +00:00
ec.balance [-c EACH_COLLECTION|<collection_name>] [-force] [-dataCenter <data_center>]
2019-05-30 08:38:59 +00:00
Algorithm:
func EcBalance() {
2019-06-11 04:32:56 +00:00
for each collection:
balanceEcVolumes(collectionName)
for each rack:
balanceEcRack(rack)
}
func balanceEcVolumes(collectionName){
for each volume:
doDeduplicateEcShards(volumeId)
tracks rack~shardCount mapping
for each volume:
doBalanceEcShardsAcrossRacks(volumeId)
for each volume:
doBalanceEcShardsWithinRacks(volumeId)
}
// spread ec shards into more racks
func doBalanceEcShardsAcrossRacks(volumeId){
tracks rack~volumeIdShardCount mapping
averageShardsPerEcRack = totalShardNumber / numRacks // totalShardNumber is 14 for now, later could varies for each dc
ecShardsToMove = select overflown ec shards from racks with ec shard counts > averageShardsPerEcRack
for each ecShardsToMove {
destRack = pickOneRack(rack~shardCount, rack~volumeIdShardCount, averageShardsPerEcRack)
destVolumeServers = volume servers on the destRack
pickOneEcNodeAndMoveOneShard(destVolumeServers)
}
}
func doBalanceEcShardsWithinRacks(volumeId){
racks = collect all racks that the volume id is on
for rack, shards := range racks
doBalanceEcShardsWithinOneRack(volumeId, shards, rack)
}
// move ec shards
func doBalanceEcShardsWithinOneRack(volumeId, shards, rackId){
tracks volumeServer~volumeIdShardCount mapping
averageShardCount = len(shards) / numVolumeServers
volumeServersOverAverage = volume servers with volumeId's ec shard counts > averageShardsPerEcRack
ecShardsToMove = select overflown ec shards from volumeServersOverAverage
for each ecShardsToMove {
destVolumeServer = pickOneVolumeServer(volumeServer~shardCount, volumeServer~volumeIdShardCount, averageShardCount)
pickOneEcNodeAndMoveOneShard(destVolumeServers)
2019-05-30 08:38:59 +00:00
}
}
2019-06-11 04:32:56 +00:00
// move ec shards while keeping shard distribution for the same volume unchanged or more even
func balanceEcRack(rack){
averageShardCount = total shards / numVolumeServers
for hasMovedOneEcShard {
sort all volume servers ordered by the number of local ec shards
pick the volume server A with the lowest number of ec shards x
pick the volume server B with the highest number of ec shards y
if y > averageShardCount and x +1 <= averageShardCount {
if B has a ec shard with volume id v that A does not have {
move one ec shard v from B to A
hasMovedOneEcShard = true
2019-05-30 08:38:59 +00:00
}
}
}
}
`
}
func (c *commandEcBalance) Do(args []string, commandEnv *CommandEnv, writer io.Writer) (err error) {
2019-05-30 08:38:59 +00:00
balanceCommand := flag.NewFlagSet(c.Name(), flag.ContinueOnError)
2019-06-03 09:26:31 +00:00
collection := balanceCommand.String("collection", "EACH_COLLECTION", "collection name, or \"EACH_COLLECTION\" for each collection")
2019-05-30 08:38:59 +00:00
dc := balanceCommand.String("dataCenter", "", "only apply the balancing for this dataCenter")
2019-06-03 09:26:31 +00:00
applyBalancing := balanceCommand.Bool("force", false, "apply the balancing plan")
2019-05-30 08:38:59 +00:00
if err = balanceCommand.Parse(args); err != nil {
return nil
}
2022-05-31 21:48:46 +00:00
infoAboutSimulationMode(writer, *applyBalancing, "-force")
2019-05-30 08:38:59 +00:00
2021-12-10 21:24:38 +00:00
if err = commandEnv.confirmIsLocked(args); err != nil {
return
}
2019-06-11 04:32:56 +00:00
// collect all ec nodes
allEcNodes, totalFreeEcSlots, err := collectEcNodes(commandEnv, *dc)
2019-05-30 08:38:59 +00:00
if err != nil {
return err
}
2019-06-11 04:32:56 +00:00
if totalFreeEcSlots < 1 {
return fmt.Errorf("no free ec shard slots. only %d left", totalFreeEcSlots)
}
2019-05-30 08:38:59 +00:00
2019-06-11 04:32:56 +00:00
racks := collectRacks(allEcNodes)
2019-05-30 08:38:59 +00:00
2019-06-11 04:32:56 +00:00
if *collection == "EACH_COLLECTION" {
collections, err := ListCollectionNames(commandEnv, false, true)
if err != nil {
return err
2019-05-30 08:38:59 +00:00
}
2019-06-11 04:32:56 +00:00
fmt.Printf("balanceEcVolumes collections %+v\n", len(collections))
for _, c := range collections {
fmt.Printf("balanceEcVolumes collection %+v\n", c)
if err = balanceEcVolumes(commandEnv, c, allEcNodes, racks, *applyBalancing); err != nil {
2019-05-30 08:38:59 +00:00
return err
}
}
2019-06-11 04:32:56 +00:00
} else {
if err = balanceEcVolumes(commandEnv, *collection, allEcNodes, racks, *applyBalancing); err != nil {
return err
}
}
2019-05-30 08:38:59 +00:00
if err := balanceEcRacks(commandEnv, racks, *applyBalancing); err != nil {
2019-06-11 04:32:56 +00:00
return fmt.Errorf("balance ec racks: %v", err)
2019-05-30 08:38:59 +00:00
}
2019-06-11 04:32:56 +00:00
2019-05-30 08:38:59 +00:00
return nil
}
2019-06-11 04:32:56 +00:00
func collectRacks(allEcNodes []*EcNode) map[RackId]*EcRack {
// collect racks info
racks := make(map[RackId]*EcRack)
for _, ecNode := range allEcNodes {
if racks[ecNode.rack] == nil {
racks[ecNode.rack] = &EcRack{
ecNodes: make(map[EcNodeId]*EcNode),
}
}
racks[ecNode.rack].ecNodes[EcNodeId(ecNode.info.Id)] = ecNode
racks[ecNode.rack].freeEcSlot += ecNode.freeEcSlot
}
return racks
}
func balanceEcVolumes(commandEnv *CommandEnv, collection string, allEcNodes []*EcNode, racks map[RackId]*EcRack, applyBalancing bool) error {
2019-05-30 08:38:59 +00:00
fmt.Printf("balanceEcVolumes %s\n", collection)
2020-02-26 06:23:59 +00:00
if err := deleteDuplicatedEcShards(commandEnv, allEcNodes, collection, applyBalancing); err != nil {
2019-06-11 04:32:56 +00:00
return fmt.Errorf("delete duplicated collection %s ec shards: %v", collection, err)
2019-05-30 08:38:59 +00:00
}
if err := balanceEcShardsAcrossRacks(commandEnv, allEcNodes, racks, collection, applyBalancing); err != nil {
2019-06-11 04:32:56 +00:00
return fmt.Errorf("balance across racks collection %s ec shards: %v", collection, err)
2019-05-30 08:38:59 +00:00
}
if err := balanceEcShardsWithinRacks(commandEnv, allEcNodes, racks, collection, applyBalancing); err != nil {
2020-10-22 05:26:41 +00:00
return fmt.Errorf("balance within racks collection %s ec shards: %v", collection, err)
2019-06-04 03:25:02 +00:00
}
2019-05-30 08:38:59 +00:00
2019-06-04 03:25:02 +00:00
return nil
}
2020-02-26 06:23:59 +00:00
func deleteDuplicatedEcShards(commandEnv *CommandEnv, allEcNodes []*EcNode, collection string, applyBalancing bool) error {
2019-06-11 04:32:56 +00:00
// vid => []ecNode
vidLocations := collectVolumeIdToEcNodes(allEcNodes, collection)
2019-06-11 04:32:56 +00:00
// deduplicate ec shards
for vid, locations := range vidLocations {
if err := doDeduplicateEcShards(commandEnv, collection, vid, locations, applyBalancing); err != nil {
2019-06-04 03:25:02 +00:00
return err
}
}
return nil
}
2019-05-30 08:38:59 +00:00
func doDeduplicateEcShards(commandEnv *CommandEnv, collection string, vid needle.VolumeId, locations []*EcNode, applyBalancing bool) error {
2019-05-30 08:38:59 +00:00
2019-06-04 03:25:02 +00:00
// check whether this volume has ecNodes that are over average
shardToLocations := make([][]*EcNode, erasure_coding.TotalShardsCount)
for _, ecNode := range locations {
shardBits := findEcVolumeShards(ecNode, vid)
for _, shardId := range shardBits.ShardIds() {
shardToLocations[shardId] = append(shardToLocations[shardId], ecNode)
}
}
for shardId, ecNodes := range shardToLocations {
if len(ecNodes) <= 1 {
continue
}
sortEcNodesByFreeslotsAscending(ecNodes)
2019-06-04 08:32:36 +00:00
fmt.Printf("ec shard %d.%d has %d copies, keeping %v\n", vid, shardId, len(ecNodes), ecNodes[0].info.Id)
2019-06-04 03:25:02 +00:00
if !applyBalancing {
continue
}
2019-06-04 08:32:36 +00:00
duplicatedShardIds := []uint32{uint32(shardId)}
2019-06-04 03:25:02 +00:00
for _, ecNode := range ecNodes[1:] {
if err := unmountEcShards(commandEnv.option.GrpcDialOption, vid, pb.NewServerAddressFromDataNode(ecNode.info), duplicatedShardIds); err != nil {
2019-06-04 03:25:02 +00:00
return err
}
if err := sourceServerDeleteEcShards(commandEnv.option.GrpcDialOption, collection, vid, pb.NewServerAddressFromDataNode(ecNode.info), duplicatedShardIds); err != nil {
2019-05-30 08:38:59 +00:00
return err
}
2019-06-11 04:32:56 +00:00
ecNode.deleteEcVolumeShards(vid, duplicatedShardIds)
2019-05-30 08:38:59 +00:00
}
}
return nil
}
func balanceEcShardsAcrossRacks(commandEnv *CommandEnv, allEcNodes []*EcNode, racks map[RackId]*EcRack, collection string, applyBalancing bool) error {
2019-06-11 04:32:56 +00:00
// collect vid => []ecNode, since previous steps can change the locations
vidLocations := collectVolumeIdToEcNodes(allEcNodes, collection)
2019-06-11 04:32:56 +00:00
// spread the ec shards evenly
for vid, locations := range vidLocations {
if err := doBalanceEcShardsAcrossRacks(commandEnv, collection, vid, locations, racks, applyBalancing); err != nil {
2019-06-11 04:32:56 +00:00
return err
}
}
return nil
}
func doBalanceEcShardsAcrossRacks(commandEnv *CommandEnv, collection string, vid needle.VolumeId, locations []*EcNode, racks map[RackId]*EcRack, applyBalancing bool) error {
2019-06-11 04:32:56 +00:00
// calculate average number of shards an ec rack should have for one volume
averageShardsPerEcRack := ceilDivide(erasure_coding.TotalShardsCount, len(racks))
// see the volume's shards are in how many racks, and how many in each rack
rackToShardCount := groupByCount(locations, func(ecNode *EcNode) (id string, count int) {
shardBits := findEcVolumeShards(ecNode, vid)
return string(ecNode.rack), shardBits.ShardIdCount()
})
rackEcNodesWithVid := groupBy(locations, func(ecNode *EcNode) string {
return string(ecNode.rack)
})
// ecShardsToMove = select overflown ec shards from racks with ec shard counts > averageShardsPerEcRack
ecShardsToMove := make(map[erasure_coding.ShardId]*EcNode)
for rackId, count := range rackToShardCount {
if count > averageShardsPerEcRack {
possibleEcNodes := rackEcNodesWithVid[rackId]
for shardId, ecNode := range pickNEcShardsToMoveFrom(possibleEcNodes, vid, count-averageShardsPerEcRack) {
ecShardsToMove[shardId] = ecNode
}
}
}
for shardId, ecNode := range ecShardsToMove {
rackId := pickOneRack(racks, rackToShardCount, averageShardsPerEcRack)
if rackId == "" {
fmt.Printf("ec shard %d.%d at %s can not find a destination rack\n", vid, shardId, ecNode.info.Id)
continue
}
2019-06-11 04:32:56 +00:00
var possibleDestinationEcNodes []*EcNode
for _, n := range racks[rackId].ecNodes {
possibleDestinationEcNodes = append(possibleDestinationEcNodes, n)
}
err := pickOneEcNodeAndMoveOneShard(commandEnv, averageShardsPerEcRack, ecNode, collection, vid, shardId, possibleDestinationEcNodes, applyBalancing)
2019-06-11 04:32:56 +00:00
if err != nil {
return err
}
rackToShardCount[string(rackId)] += 1
rackToShardCount[string(ecNode.rack)] -= 1
racks[rackId].freeEcSlot -= 1
racks[ecNode.rack].freeEcSlot += 1
}
return nil
}
func pickOneRack(rackToEcNodes map[RackId]*EcRack, rackToShardCount map[string]int, averageShardsPerEcRack int) RackId {
// TODO later may need to add some randomness
for rackId, rack := range rackToEcNodes {
if rackToShardCount[string(rackId)] >= averageShardsPerEcRack {
continue
}
if rack.freeEcSlot <= 0 {
continue
}
return rackId
}
return ""
}
func balanceEcShardsWithinRacks(commandEnv *CommandEnv, allEcNodes []*EcNode, racks map[RackId]*EcRack, collection string, applyBalancing bool) error {
2019-06-11 04:32:56 +00:00
// collect vid => []ecNode, since previous steps can change the locations
vidLocations := collectVolumeIdToEcNodes(allEcNodes, collection)
2019-06-11 04:32:56 +00:00
// spread the ec shards evenly
for vid, locations := range vidLocations {
// see the volume's shards are in how many racks, and how many in each rack
rackToShardCount := groupByCount(locations, func(ecNode *EcNode) (id string, count int) {
shardBits := findEcVolumeShards(ecNode, vid)
return string(ecNode.rack), shardBits.ShardIdCount()
})
rackEcNodesWithVid := groupBy(locations, func(ecNode *EcNode) string {
return string(ecNode.rack)
})
for rackId, _ := range rackToShardCount {
var possibleDestinationEcNodes []*EcNode
for _, n := range racks[RackId(rackId)].ecNodes {
2021-02-16 14:22:50 +00:00
if _, found := n.info.DiskInfos[string(types.HardDriveType)]; found {
possibleDestinationEcNodes = append(possibleDestinationEcNodes, n)
}
2019-06-11 04:32:56 +00:00
}
sourceEcNodes := rackEcNodesWithVid[rackId]
averageShardsPerEcNode := ceilDivide(rackToShardCount[rackId], len(possibleDestinationEcNodes))
if err := doBalanceEcShardsWithinOneRack(commandEnv, averageShardsPerEcNode, collection, vid, sourceEcNodes, possibleDestinationEcNodes, applyBalancing); err != nil {
2019-06-11 04:32:56 +00:00
return err
}
}
}
return nil
}
func doBalanceEcShardsWithinOneRack(commandEnv *CommandEnv, averageShardsPerEcNode int, collection string, vid needle.VolumeId, existingLocations, possibleDestinationEcNodes []*EcNode, applyBalancing bool) error {
2019-05-30 08:38:59 +00:00
for _, ecNode := range existingLocations {
shardBits := findEcVolumeShards(ecNode, vid)
overLimitCount := shardBits.ShardIdCount() - averageShardsPerEcNode
for _, shardId := range shardBits.ShardIds() {
if overLimitCount <= 0 {
break
}
fmt.Printf("%s has %d overlimit, moving ec shard %d.%d\n", ecNode.info.Id, overLimitCount, vid, shardId)
err := pickOneEcNodeAndMoveOneShard(commandEnv, averageShardsPerEcNode, ecNode, collection, vid, shardId, possibleDestinationEcNodes, applyBalancing)
2019-05-30 08:38:59 +00:00
if err != nil {
return err
}
overLimitCount--
}
}
return nil
}
func balanceEcRacks(commandEnv *CommandEnv, racks map[RackId]*EcRack, applyBalancing bool) error {
2019-06-11 04:32:56 +00:00
// balance one rack for all ec shards
for _, ecRack := range racks {
if err := doBalanceEcRack(commandEnv, ecRack, applyBalancing); err != nil {
2019-06-11 04:32:56 +00:00
return err
}
}
return nil
}
func doBalanceEcRack(commandEnv *CommandEnv, ecRack *EcRack, applyBalancing bool) error {
2019-06-11 04:32:56 +00:00
if len(ecRack.ecNodes) <= 1 {
return nil
}
var rackEcNodes []*EcNode
for _, node := range ecRack.ecNodes {
rackEcNodes = append(rackEcNodes, node)
}
2021-02-16 10:47:02 +00:00
ecNodeIdToShardCount := groupByCount(rackEcNodes, func(ecNode *EcNode) (id string, count int) {
2021-02-16 13:13:48 +00:00
diskInfo, found := ecNode.info.DiskInfos[string(types.HardDriveType)]
if !found {
return
}
2021-02-16 10:47:02 +00:00
for _, ecShardInfo := range diskInfo.EcShardInfos {
2019-06-11 04:32:56 +00:00
count += erasure_coding.ShardBits(ecShardInfo.EcIndexBits).ShardIdCount()
}
2021-02-16 10:47:02 +00:00
return ecNode.info.Id, count
2019-06-11 04:32:56 +00:00
})
var totalShardCount int
for _, count := range ecNodeIdToShardCount {
totalShardCount += count
}
averageShardCount := ceilDivide(totalShardCount, len(rackEcNodes))
hasMove := true
for hasMove {
hasMove = false
slices.SortFunc(rackEcNodes, func(a, b *EcNode) int {
return b.freeEcSlot - a.freeEcSlot
2019-06-11 04:32:56 +00:00
})
emptyNode, fullNode := rackEcNodes[0], rackEcNodes[len(rackEcNodes)-1]
emptyNodeShardCount, fullNodeShardCount := ecNodeIdToShardCount[emptyNode.info.Id], ecNodeIdToShardCount[fullNode.info.Id]
if fullNodeShardCount > averageShardCount && emptyNodeShardCount+1 <= averageShardCount {
emptyNodeIds := make(map[uint32]bool)
2021-02-16 13:13:48 +00:00
if emptyDiskInfo, found := emptyNode.info.DiskInfos[string(types.HardDriveType)]; found {
for _, shards := range emptyDiskInfo.EcShardInfos {
emptyNodeIds[shards.Id] = true
}
2019-06-11 04:32:56 +00:00
}
2021-02-16 13:13:48 +00:00
if fullDiskInfo, found := fullNode.info.DiskInfos[string(types.HardDriveType)]; found {
for _, shards := range fullDiskInfo.EcShardInfos {
if _, found := emptyNodeIds[shards.Id]; !found {
for _, shardId := range erasure_coding.ShardBits(shards.EcIndexBits).ShardIds() {
fmt.Printf("%s moves ec shards %d.%d to %s\n", fullNode.info.Id, shards.Id, shardId, emptyNode.info.Id)
err := moveMountedShardToEcNode(commandEnv, fullNode, shards.Collection, needle.VolumeId(shards.Id), shardId, emptyNode, applyBalancing)
if err != nil {
return err
}
ecNodeIdToShardCount[emptyNode.info.Id]++
ecNodeIdToShardCount[fullNode.info.Id]--
hasMove = true
break
2019-06-11 04:32:56 +00:00
}
break
}
}
}
}
}
return nil
}
func pickOneEcNodeAndMoveOneShard(commandEnv *CommandEnv, averageShardsPerEcNode int, existingLocation *EcNode, collection string, vid needle.VolumeId, shardId erasure_coding.ShardId, possibleDestinationEcNodes []*EcNode, applyBalancing bool) error {
2019-05-30 08:38:59 +00:00
sortEcNodesByFreeslotsDescending(possibleDestinationEcNodes)
2019-05-30 08:38:59 +00:00
for _, destEcNode := range possibleDestinationEcNodes {
if destEcNode.info.Id == existingLocation.info.Id {
continue
}
if destEcNode.freeEcSlot <= 0 {
continue
}
if findEcVolumeShards(destEcNode, vid).ShardIdCount() >= averageShardsPerEcNode {
continue
}
fmt.Printf("%s moves ec shard %d.%d to %s\n", existingLocation.info.Id, vid, shardId, destEcNode.info.Id)
err := moveMountedShardToEcNode(commandEnv, existingLocation, collection, vid, shardId, destEcNode, applyBalancing)
2019-05-30 08:38:59 +00:00
if err != nil {
return err
}
return nil
}
return nil
}
2019-06-11 04:33:32 +00:00
func pickNEcShardsToMoveFrom(ecNodes []*EcNode, vid needle.VolumeId, n int) map[erasure_coding.ShardId]*EcNode {
2019-06-11 04:32:56 +00:00
picked := make(map[erasure_coding.ShardId]*EcNode)
var candidateEcNodes []*CandidateEcNode
for _, ecNode := range ecNodes {
shardBits := findEcVolumeShards(ecNode, vid)
if shardBits.ShardIdCount() > 0 {
candidateEcNodes = append(candidateEcNodes, &CandidateEcNode{
ecNode: ecNode,
shardCount: shardBits.ShardIdCount(),
})
2019-05-30 08:38:59 +00:00
}
}
slices.SortFunc(candidateEcNodes, func(a, b *CandidateEcNode) int {
return b.shardCount - a.shardCount
2019-06-11 04:32:56 +00:00
})
for i := 0; i < n; i++ {
selectedEcNodeIndex := -1
for i, candidateEcNode := range candidateEcNodes {
shardBits := findEcVolumeShards(candidateEcNode.ecNode, vid)
if shardBits > 0 {
selectedEcNodeIndex = i
for _, shardId := range shardBits.ShardIds() {
candidateEcNode.shardCount--
picked[shardId] = candidateEcNode.ecNode
candidateEcNode.ecNode.deleteEcVolumeShards(vid, []uint32{uint32(shardId)})
break
}
break
2019-06-06 06:20:26 +00:00
}
}
2019-06-11 04:32:56 +00:00
if selectedEcNodeIndex >= 0 {
ensureSortedEcNodes(candidateEcNodes, selectedEcNodeIndex, func(i, j int) bool {
return candidateEcNodes[i].shardCount > candidateEcNodes[j].shardCount
})
}
2019-06-06 06:20:26 +00:00
2019-06-11 04:32:56 +00:00
}
return picked
2019-06-06 06:20:26 +00:00
}
func collectVolumeIdToEcNodes(allEcNodes []*EcNode, collection string) map[needle.VolumeId][]*EcNode {
2019-06-11 04:32:56 +00:00
vidLocations := make(map[needle.VolumeId][]*EcNode)
for _, ecNode := range allEcNodes {
2021-02-16 13:13:48 +00:00
diskInfo, found := ecNode.info.DiskInfos[string(types.HardDriveType)]
if !found {
continue
}
2021-02-16 10:47:02 +00:00
for _, shardInfo := range diskInfo.EcShardInfos {
// ignore if not in current collection
if shardInfo.Collection == collection {
vidLocations[needle.VolumeId(shardInfo.Id)] = append(vidLocations[needle.VolumeId(shardInfo.Id)], ecNode)
}
2019-06-06 06:20:26 +00:00
}
}
2019-06-11 04:32:56 +00:00
return vidLocations
2019-06-06 06:20:26 +00:00
}