lib/config, lib/model: Auto adjust pullers based on desired amount of pending data (#4748)

This makes the number of pullers vary with the desired amount of outstanding requests instead of being a fixed number.
This commit is contained in:
Jakob Borg 2018-02-25 10:14:02 +01:00 committed by GitHub
parent 158859a1e2
commit 42cc64e2ed
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
9 changed files with 175 additions and 84 deletions

View File

@ -32,7 +32,7 @@ import (
const (
OldestHandledVersion = 10
CurrentVersion = 26
CurrentVersion = 27
MaxRescanIntervalS = 365 * 24 * 60 * 60
)
@ -312,6 +312,9 @@ func (cfg *Configuration) clean() error {
if cfg.Version == 25 {
convertV25V26(cfg)
}
if cfg.Version == 26 {
convertV26V27(cfg)
}
// Build a list of available devices
existingDevices := make(map[protocol.DeviceID]bool)
@ -371,6 +374,17 @@ func (cfg *Configuration) clean() error {
return nil
}
func convertV26V27(cfg *Configuration) {
for i := range cfg.Folders {
f := &cfg.Folders[i]
if f.DeprecatedPullers != 0 {
f.PullerMaxPendingKiB = 128 * f.DeprecatedPullers
f.DeprecatedPullers = 0
}
}
cfg.Version = 27
}
func convertV25V26(cfg *Configuration) {
// triggers database update
cfg.Version = 26

View File

@ -107,7 +107,6 @@ func TestDeviceConfig(t *testing.T) {
FSWatcherEnabled: false,
FSWatcherDelayS: 10,
Copiers: 0,
Pullers: 0,
Hashers: 0,
AutoNormalize: true,
MinDiskFree: Size{1, "%"},

View File

@ -40,7 +40,7 @@ type FolderConfiguration struct {
MinDiskFree Size `xml:"minDiskFree" json:"minDiskFree"`
Versioning VersioningConfiguration `xml:"versioning" json:"versioning"`
Copiers int `xml:"copiers" json:"copiers"` // This defines how many files are handled concurrently.
Pullers int `xml:"pullers" json:"pullers"` // Defines how many blocks are fetched at the same time, possibly between separate copier routines.
PullerMaxPendingKiB int `xml:"pullerMaxPendingKiB" json:"pullerMaxPendingKiB"`
Hashers int `xml:"hashers" json:"hashers"` // Less than one sets the value to the number of cores. These are CPU bound due to hashing.
Order PullOrder `xml:"order" json:"order"`
IgnoreDelete bool `xml:"ignoreDelete" json:"ignoreDelete"`
@ -57,6 +57,7 @@ type FolderConfiguration struct {
DeprecatedReadOnly bool `xml:"ro,attr,omitempty" json:"-"`
DeprecatedMinDiskFreePct float64 `xml:"minDiskFreePct,omitempty" json:"-"`
DeprecatedPullers int `xml:"pullers,omitempty" json:"-"`
}
type FolderDeviceConfiguration struct {

16
lib/config/testdata/v27.xml vendored Normal file
View File

@ -0,0 +1,16 @@
<configuration version="27">
<folder id="test" path="testdata" type="readonly" ignorePerms="false" rescanIntervalS="600" fsNotifications="false" notifyDelayS="10" autoNormalize="true">
<filesystemType>basic</filesystemType>
<device id="AIR6LPZ-7K4PTTV-UXQSMUU-CPQ5YWH-OEDFIIQ-JUG777G-2YQXXR5-YD6AWQR"></device>
<device id="P56IOI7-MZJNU2Y-IQGDREY-DM2MGTI-MGL3BXN-PQ6W5BM-TBBZ4TJ-XZWICQ2"></device>
<minDiskFree unit="%">1</minDiskFree>
<maxConflicts>-1</maxConflicts>
<fsync>true</fsync>
</folder>
<device id="AIR6LPZ-7K4PTTV-UXQSMUU-CPQ5YWH-OEDFIIQ-JUG777G-2YQXXR5-YD6AWQR" name="node one" compression="metadata">
<address>tcp://a</address>
</device>
<device id="P56IOI7-MZJNU2Y-IQGDREY-DM2MGTI-MGL3BXN-PQ6W5BM-TBBZ4TJ-XZWICQ2" name="node two" compression="metadata">
<address>tcp://b</address>
</device>
</configuration>

View File

@ -15,6 +15,7 @@ import (
"runtime"
"sort"
"strings"
stdsync "sync"
"time"
"github.com/syncthing/syncthing/lib/config"
@ -79,8 +80,9 @@ const (
const (
defaultCopiers = 2
defaultPullers = 64
defaultPullerPause = 60 * time.Second
defaultPullerPendingKiB = 8192 // must be larger than block size
maxPullerIterations = 3
)
@ -114,20 +116,23 @@ func newSendReceiveFolder(model *Model, cfg config.FolderConfiguration, ver vers
errorsMut: sync.NewMutex(),
}
f.configureCopiersAndPullers()
return f
}
func (f *sendReceiveFolder) configureCopiersAndPullers() {
if f.Copiers == 0 {
f.Copiers = defaultCopiers
}
if f.Pullers == 0 {
f.Pullers = defaultPullers
// If the configured max amount of pending data is zero, we use the
// default. If it's configured to something non-zero but less than the
// protocol block size we adjust it upwards accordingly.
if f.PullerMaxPendingKiB == 0 {
f.PullerMaxPendingKiB = defaultPullerPendingKiB
}
if blockSizeKiB := protocol.BlockSize / 1024; f.PullerMaxPendingKiB < blockSizeKiB {
f.PullerMaxPendingKiB = blockSizeKiB
}
f.pause = f.basePause()
return f
}
// Serve will run scans and pulls. It will return when Stop()ed or on a
@ -317,7 +322,7 @@ func (f *sendReceiveFolder) pullerIteration(ignores *ignore.Matcher, ignoresChan
doneWg := sync.NewWaitGroup()
updateWg := sync.NewWaitGroup()
l.Debugln(f, "c", f.Copiers, "p", f.Pullers)
l.Debugln(f, "copiers:", f.Copiers, "pullerPendingKiB:", f.PullerMaxPendingKiB)
updateWg.Add(1)
go func() {
@ -335,14 +340,12 @@ func (f *sendReceiveFolder) pullerIteration(ignores *ignore.Matcher, ignoresChan
}()
}
for i := 0; i < f.Pullers; i++ {
pullWg.Add(1)
go func() {
// pullerRoutine finishes when pullChan is closed
f.pullerRoutine(pullChan, finisherChan)
pullWg.Done()
}()
}
doneWg.Add(1)
// finisherRoutine finishes when finisherChan is closed
@ -1340,19 +1343,43 @@ func verifyBuffer(buf []byte, block protocol.BlockInfo) ([]byte, error) {
}
func (f *sendReceiveFolder) pullerRoutine(in <-chan pullBlockState, out chan<- *sharedPullerState) {
requestLimiter := newByteSemaphore(f.PullerMaxPendingKiB * 1024)
wg := sync.NewWaitGroup()
for state := range in {
if state.failed() != nil {
out <- state.sharedPullerState
continue
}
// The requestLimiter limits how many pending block requests we have
// ongoing at any given time, based on the size of the blocks
// themselves.
state := state
bytes := int(state.block.Size)
requestLimiter.take(bytes)
wg.Add(1)
go func() {
defer wg.Done()
defer requestLimiter.give(bytes)
f.pullBlock(state, out)
}()
}
wg.Wait()
}
func (f *sendReceiveFolder) pullBlock(state pullBlockState, out chan<- *sharedPullerState) {
// Get an fd to the temporary file. Technically we don't need it until
// after fetching the block, but if we run into an error here there is
// no point in issuing the request to the network.
fd, err := state.tempFile()
if err != nil {
out <- state.sharedPullerState
continue
return
}
if !f.DisableSparseFiles && state.reused == 0 && state.block.IsEmpty() {
@ -1360,7 +1387,7 @@ func (f *sendReceiveFolder) pullerRoutine(in <-chan pullBlockState, out chan<- *
// requested it and handled it correctly.
state.pullDone(state.block)
out <- state.sharedPullerState
continue
return
}
var lastError error
@ -1409,7 +1436,6 @@ func (f *sendReceiveFolder) pullerRoutine(in <-chan pullBlockState, out chan<- *
break
}
out <- state.sharedPullerState
}
}
func (f *sendReceiveFolder) performFinish(ignores *ignore.Matcher, state *sharedPullerState, dbUpdateChan chan<- dbUpdateJob, scanChan chan<- string) error {
@ -1899,3 +1925,41 @@ func componentCount(name string) int {
}
return count
}
type byteSemaphore struct {
max int
available int
mut stdsync.Mutex
cond *stdsync.Cond
}
func newByteSemaphore(max int) *byteSemaphore {
s := byteSemaphore{
max: max,
available: max,
}
s.cond = stdsync.NewCond(&s.mut)
return &s
}
func (s *byteSemaphore) take(bytes int) {
if bytes > s.max {
panic("bug: more than max bytes will never be available")
}
s.mut.Lock()
for bytes > s.available {
s.cond.Wait()
}
s.available -= bytes
s.mut.Unlock()
}
func (s *byteSemaphore) give(bytes int) {
s.mut.Lock()
if s.available+bytes > s.max {
panic("bug: can never give more than max")
}
s.available += bytes
s.cond.Broadcast()
s.mut.Unlock()
}

View File

@ -17,6 +17,7 @@ import (
"testing"
"time"
"github.com/syncthing/syncthing/lib/config"
"github.com/syncthing/syncthing/lib/db"
"github.com/syncthing/syncthing/lib/fs"
"github.com/syncthing/syncthing/lib/ignore"
@ -107,6 +108,9 @@ func setUpSendReceiveFolder(model *Model) *sendReceiveFolder {
model: model,
initialScanFinished: make(chan struct{}),
ctx: context.TODO(),
FolderConfiguration: config.FolderConfiguration{
PullerMaxPendingKiB: defaultPullerPendingKiB,
},
},
fs: fs.NewMtimeFS(fs.NewFilesystem(fs.FilesystemTypeBasic, "testdata"), db.NewNamespacedKV(model.db, "mtime")),

View File

@ -8,7 +8,6 @@
<minDiskFree unit="%">1</minDiskFree>
<versioning></versioning>
<copiers>1</copiers>
<pullers>16</pullers>
<hashers>0</hashers>
<order>random</order>
<ignoreDelete>false</ignoreDelete>
@ -28,7 +27,6 @@
<minDiskFree unit="%">1</minDiskFree>
<versioning></versioning>
<copiers>1</copiers>
<pullers>16</pullers>
<hashers>0</hashers>
<order>random</order>
<ignoreDelete>false</ignoreDelete>

View File

@ -7,7 +7,6 @@
<minDiskFree unit="%">1</minDiskFree>
<versioning></versioning>
<copiers>1</copiers>
<pullers>16</pullers>
<hashers>0</hashers>
<order>random</order>
<ignoreDelete>false</ignoreDelete>
@ -27,7 +26,6 @@
<minDiskFree unit="%">1</minDiskFree>
<versioning></versioning>
<copiers>1</copiers>
<pullers>16</pullers>
<hashers>0</hashers>
<order>random</order>
<ignoreDelete>false</ignoreDelete>
@ -47,7 +45,6 @@
<minDiskFree unit="%">1</minDiskFree>
<versioning></versioning>
<copiers>1</copiers>
<pullers>16</pullers>
<hashers>0</hashers>
<order>random</order>
<ignoreDelete>false</ignoreDelete>

View File

@ -9,7 +9,6 @@
<param key="keep" val="5"></param>
</versioning>
<copiers>1</copiers>
<pullers>16</pullers>
<hashers>0</hashers>
<order>random</order>
<ignoreDelete>false</ignoreDelete>
@ -29,7 +28,6 @@
<minDiskFree unit="%">1</minDiskFree>
<versioning></versioning>
<copiers>1</copiers>
<pullers>16</pullers>
<hashers>0</hashers>
<order>random</order>
<ignoreDelete>false</ignoreDelete>