allocate ec shards to volume servers

This commit is contained in:
Chris Lu 2019-05-25 02:02:44 -07:00
parent 6f4b09b6a4
commit f0e6574d5e
6 changed files with 172 additions and 14 deletions

View file

@ -5,10 +5,15 @@ import (
"flag"
"fmt"
"io"
"sort"
"sync"
"github.com/chrislusf/seaweedfs/weed/operation"
"github.com/chrislusf/seaweedfs/weed/pb/master_pb"
"github.com/chrislusf/seaweedfs/weed/pb/volume_server_pb"
"github.com/chrislusf/seaweedfs/weed/storage/erasure_coding"
"github.com/chrislusf/seaweedfs/weed/storage/needle"
"github.com/chrislusf/seaweedfs/weed/wdclient"
"google.golang.org/grpc"
)
@ -53,18 +58,28 @@ func (c *commandEcEncode) Do(args []string, commandEnv *commandEnv, writer io.Wr
ctx := context.Background()
// find volume location
locations := commandEnv.masterClient.GetLocations(uint32(*volumeId))
if len(locations) == 0 {
return fmt.Errorf("volume %d not found", *volumeId)
}
err = generateEcSlices(ctx, commandEnv.option.GrpcDialOption, needle.VolumeId(*volumeId), locations[0].Url)
// generate ec shards
err = generateEcShards(ctx, commandEnv.option.GrpcDialOption, needle.VolumeId(*volumeId), locations[0].Url)
if err != nil {
return fmt.Errorf("generate ec shards for volume %d on %s: %v", *volumeId, locations[0].Url, err)
}
// balance the ec shards to current cluster
err = balanceEcShards(ctx, commandEnv, needle.VolumeId(*volumeId), locations[0])
if err != nil {
return fmt.Errorf("balance ec shards for volume %d on %s: %v", *volumeId, locations[0].Url, err)
}
return err
}
func generateEcSlices(ctx context.Context, grpcDialOption grpc.DialOption, volumeId needle.VolumeId, sourceVolumeServer string) error {
func generateEcShards(ctx context.Context, grpcDialOption grpc.DialOption, volumeId needle.VolumeId, sourceVolumeServer string) error {
err := operation.WithVolumeServerClient(sourceVolumeServer, grpcDialOption, func(volumeServerClient volume_server_pb.VolumeServerClient) error {
_, genErr := volumeServerClient.VolumeEcGenerateSlices(ctx, &volume_server_pb.VolumeEcGenerateSlicesRequest{
@ -76,3 +91,136 @@ func generateEcSlices(ctx context.Context, grpcDialOption grpc.DialOption, volum
return err
}
func balanceEcShards(ctx context.Context, commandEnv *commandEnv, volumeId needle.VolumeId, existingLocation wdclient.Location) (err error) {
// list all possible locations
var resp *master_pb.VolumeListResponse
err = commandEnv.masterClient.WithClient(ctx, func(client master_pb.SeaweedClient) error {
resp, err = client.VolumeList(ctx, &master_pb.VolumeListRequest{})
return err
})
if err != nil {
return err
}
// find out all volume servers with one volume slot left.
var allDataNodes []*master_pb.DataNodeInfo
var totalFreeEcSlots int
eachDataNode(resp.TopologyInfo, func(dn *master_pb.DataNodeInfo) {
if freeEcSlots := countFreeShardSlots(dn); freeEcSlots > 0 {
allDataNodes = append(allDataNodes, dn)
totalFreeEcSlots += freeEcSlots
}
})
if totalFreeEcSlots < erasure_coding.TotalShardsCount {
return fmt.Errorf("not enough free ec shard slots. only %d left", totalFreeEcSlots)
}
sort.Slice(allDataNodes, func(i, j int) bool {
return countFreeShardSlots(allDataNodes[j]) < countFreeShardSlots(allDataNodes[i])
})
if len(allDataNodes) > erasure_coding.TotalShardsCount {
allDataNodes = allDataNodes[:erasure_coding.TotalShardsCount]
}
// calculate how many shards to allocate for these servers
allocated := balancedEcDistribution(allDataNodes)
// ask the data nodes to copy from the source volume server
err = parallelCopyEcShardsFromSource(ctx, commandEnv.option.GrpcDialOption, allDataNodes, allocated, volumeId, existingLocation)
if err != nil {
return nil
}
// ask the source volume server to clean up copied ec shards
// ask the source volume server to delete the original volume
return err
}
func parallelCopyEcShardsFromSource(ctx context.Context, grpcDialOption grpc.DialOption,
targetServers []*master_pb.DataNodeInfo, allocated []int,
volumeId needle.VolumeId, existingLocation wdclient.Location) (err error) {
// parallelize
var wg sync.WaitGroup
startFromShardId := 0
for i, server := range targetServers {
if allocated[i] <= 0 {
continue
}
wg.Add(1)
go func(server *master_pb.DataNodeInfo, startFromShardId int, shardCount int) {
defer wg.Done()
copyErr := oneServerCopyEcShardsFromSource(ctx, grpcDialOption, server, startFromShardId, shardCount, volumeId, existingLocation)
if copyErr != nil {
err = copyErr
}
}(server, startFromShardId, allocated[i])
startFromShardId += allocated[i]
}
wg.Wait()
return err
}
func oneServerCopyEcShardsFromSource(ctx context.Context, grpcDialOption grpc.DialOption,
targetServer *master_pb.DataNodeInfo, startFromShardId int, shardCount int,
volumeId needle.VolumeId, existingLocation wdclient.Location) (err error) {
if targetServer.Id == existingLocation.Url {
return nil
}
for shardId := startFromShardId; shardId < startFromShardId+shardCount; shardId++ {
fmt.Printf("copy %d.%d %s => %s\n", volumeId, shardId, existingLocation.Url, targetServer.Id)
}
return nil
}
func balancedEcDistribution(servers []*master_pb.DataNodeInfo) (allocated []int) {
freeSlots := make([]int, len(servers))
allocated = make([]int, len(servers))
for i, server := range servers {
freeSlots[i] = countFreeShardSlots(server)
}
allocatedCount := 0
for allocatedCount < erasure_coding.TotalShardsCount {
for i, _ := range servers {
if freeSlots[i]-allocated[i] > 0 {
allocated[i] += 1
allocatedCount += 1
}
if allocatedCount >= erasure_coding.TotalShardsCount {
break
}
}
}
return allocated
}
func eachDataNode(topo *master_pb.TopologyInfo, fn func(*master_pb.DataNodeInfo)) {
for _, dc := range topo.DataCenterInfos {
for _, rack := range dc.RackInfos {
for _, dn := range rack.DataNodeInfos {
fn(dn)
}
}
}
}
func countShards(ecShardInfos []*master_pb.VolumeEcShardInformationMessage) (count int) {
for _, ecShardInfo := range ecShardInfos {
shardBits := erasure_coding.ShardBits(ecShardInfo.EcIndexBits)
count += shardBits.ShardIdCount()
}
return
}
func countFreeShardSlots(dn *master_pb.DataNodeInfo) (count int) {
return int(dn.FreeVolumeCount)*10 - countShards(dn.EcShardInfos)
}

View file

@ -15,6 +15,7 @@ import (
const (
DataShardsCount = 10
ParityShardsCount = 4
TotalShardsCount = DataShardsCount + ParityShardsCount
ErasureCodingLargeBlockSize = 1024 * 1024 * 1024 // 1GB
ErasureCodingSmallBlockSize = 1024 * 1024 // 1MB
)
@ -93,7 +94,7 @@ func encodeData(file *os.File, enc reedsolomon.Encoder, startOffset, blockSize i
}
func openEcFiles(baseFileName string, forRead bool) (files []*os.File, err error) {
for i := 0; i < DataShardsCount+ParityShardsCount; i++ {
for i := 0; i < TotalShardsCount; i++ {
fname := baseFileName + ToExt(i)
openOption := os.O_TRUNC | os.O_CREATE | os.O_WRONLY
if forRead {
@ -138,7 +139,7 @@ func encodeDataOneBatch(file *os.File, enc reedsolomon.Encoder, startOffset, blo
return err
}
for i := 0; i < DataShardsCount+ParityShardsCount; i++ {
for i := 0; i < TotalShardsCount; i++ {
_, err := outputs[i].Write(buffers[i])
if err != nil {
return err
@ -154,7 +155,7 @@ func encodeDatFile(remainingSize int64, err error, baseFileName string, bufferSi
if err != nil {
return fmt.Errorf("failed to create encoder: %v", err)
}
buffers := make([][]byte, DataShardsCount+ParityShardsCount)
buffers := make([][]byte, TotalShardsCount)
outputs, err := openEcFiles(baseFileName, false)
defer closeEcFiles(outputs)
if err != nil {

View file

@ -153,9 +153,9 @@ func readFromOtherEcFiles(ecFiles []*os.File, ecFileIndex int, ecFileOffset int6
return nil, fmt.Errorf("failed to create encoder: %v", err)
}
bufs := make([][]byte, DataShardsCount+ParityShardsCount)
bufs := make([][]byte, TotalShardsCount)
for i := 0; i < DataShardsCount; {
n := int(rand.Int31n(DataShardsCount + ParityShardsCount))
n := int(rand.Int31n(TotalShardsCount))
if n == ecFileIndex || bufs[n] != nil {
continue
}

View file

@ -36,6 +36,10 @@ func (ecInfo *EcVolumeInfo) ShardIds() (ret []ShardId) {
return ecInfo.ShardBits.ShardIds()
}
func (ecInfo *EcVolumeInfo) ShardIdCount() (count int) {
return ecInfo.ShardBits.ShardIdCount()
}
func (ecInfo *EcVolumeInfo) Minus(other *EcVolumeInfo) (*EcVolumeInfo) {
ret := &EcVolumeInfo{
VolumeId: ecInfo.VolumeId,
@ -69,7 +73,7 @@ func (b ShardBits) HasShardId(id ShardId) bool {
}
func (b ShardBits) ShardIds() (ret []ShardId) {
for i := ShardId(0); i < DataShardsCount+ParityShardsCount; i++ {
for i := ShardId(0); i < TotalShardsCount; i++ {
if b.HasShardId(i) {
ret = append(ret, i)
}
@ -77,6 +81,13 @@ func (b ShardBits) ShardIds() (ret []ShardId) {
return
}
func (b ShardBits) ShardIdCount() (count int) {
for count = 0; b > 0; count++ {
b &= b - 1
}
return
}
func (b ShardBits) Minus(other ShardBits) (ShardBits) {
return b &^ other
}

View file

@ -30,11 +30,11 @@ func (dn *DataNode) UpdateEcShards(actualShards []*erasure_coding.EcVolumeInfo)
} else {
// found, but maybe the actual shard could be missing
a := actualEcShards.Minus(ecShards)
if len(a.ShardIds()) > 0 {
if a.ShardIdCount() > 0 {
newShards = append(newShards, a)
}
d := ecShards.Minus(actualEcShards)
if len(d.ShardIds()) > 0 {
if d.ShardIdCount() > 0 {
deletedShards = append(deletedShards, d)
}
}

View file

@ -7,11 +7,9 @@ import (
"github.com/chrislusf/seaweedfs/weed/storage/needle"
)
const shardCount = erasure_coding.DataShardsCount + erasure_coding.ParityShardsCount
type EcShardLocations struct {
Collection string
locations [shardCount][]*DataNode
locations [erasure_coding.TotalShardsCount][]*DataNode
}
func (t *Topology) SyncDataNodeEcShards(shardInfos []*master_pb.VolumeEcShardInformationMessage, dn *DataNode) (newShards, deletedShards []*erasure_coding.EcVolumeInfo) {