"fmt"
"log"
"math"
- "os"
"runtime"
"sort"
"strings"
"git.curoverse.com/arvados.git/sdk/go/arvados"
"git.curoverse.com/arvados.git/sdk/go/keepclient"
+ "github.com/sirupsen/logrus"
)
-// CheckConfig returns an error if anything is wrong with the given
-// config and runOptions.
-func CheckConfig(config Config, runOptions RunOptions) error {
- if len(config.KeepServiceList.Items) > 0 && config.KeepServiceTypes != nil {
- return fmt.Errorf("cannot specify both KeepServiceList and KeepServiceTypes in config")
- }
- if !runOptions.Once && config.RunPeriod == arvados.Duration(0) {
- return fmt.Errorf("you must either use the -once flag, or specify RunPeriod in config")
- }
- return nil
-}
-
// Balancer compares the contents of keepstore servers with the
// collections stored in Arvados, and issues pull/trash requests
// needed to get (closer to) the optimal data layout.
// BlobSignatureTTL; and all N existing replicas of a given data block
// are in the N best positions in rendezvous probe order.
type Balancer struct {
+ Logger *logrus.Logger
+ Dumper *logrus.Logger
+ Metrics *metrics
+
*BlockStateMap
KeepServices map[string]*KeepService
DefaultReplication int
- Logger *log.Logger
- Dumper *log.Logger
MinMtime int64
classes []string
func (bal *Balancer) Run(config Config, runOptions RunOptions) (nextRunOptions RunOptions, err error) {
nextRunOptions = runOptions
- bal.Dumper = runOptions.Dumper
- bal.Logger = runOptions.Logger
- if bal.Logger == nil {
- bal.Logger = log.New(os.Stderr, "", log.LstdFlags)
- }
-
- defer timeMe(bal.Logger, "Run")()
+ defer bal.time("sweep", "wall clock time to run one full sweep")()
if len(config.KeepServiceList.Items) > 0 {
err = bal.SetKeepServices(config.KeepServiceList)
return
}
}
- bal.dedupDevices()
+ bal.cleanupMounts()
if err = bal.CheckSanityEarly(&config.Client); err != nil {
return
})
}
-func (bal *Balancer) dedupDevices() {
+func (bal *Balancer) cleanupMounts() {
rwdev := map[string]*KeepService{}
for _, srv := range bal.KeepServices {
for _, mnt := range srv.mounts {
}
srv.mounts = dedup
}
+ for _, srv := range bal.KeepServices {
+ for _, mnt := range srv.mounts {
+ if mnt.Replication <= 0 {
+ log.Printf("%s: mount %s reports replication=%d, using replication=1", srv, mnt.UUID, mnt.Replication)
+ mnt.Replication = 1
+ }
+ }
+ }
}
// CheckSanityEarly checks for configuration and runtime errors that
//
// It encodes the resulting information in BlockStateMap.
func (bal *Balancer) GetCurrentState(c *arvados.Client, pageSize, bufs int) error {
- defer timeMe(bal.Logger, "GetCurrentState")()
+ defer bal.time("get_state", "wall clock time to get current state")()
bal.BlockStateMap = NewBlockStateMap()
dd, err := c.DiscoveryDocument()
bal.DefaultReplication = dd.DefaultCollectionReplication
bal.MinMtime = time.Now().UnixNano() - dd.BlobSignatureTTL*1e9
- errs := make(chan error, 2+len(bal.KeepServices))
+ errs := make(chan error, 1)
wg := sync.WaitGroup{}
// When a device is mounted more than once, we will get its
bal.logf("mount %s: retrieve index from %s", mounts[0], mounts[0].KeepService)
idx, err := mounts[0].KeepService.IndexMount(c, mounts[0].UUID, "")
if err != nil {
- errs <- fmt.Errorf("%s: retrieve index: %v", mounts[0], err)
+ select {
+ case errs <- fmt.Errorf("%s: retrieve index: %v", mounts[0], err):
+ default:
+ }
return
}
if len(errs) > 0 {
return
}
for _, mount := range mounts {
- bal.logf("%s: add %d replicas to map", mount, len(idx))
+ bal.logf("%s: add %d entries to map", mount, len(idx))
bal.BlockStateMap.AddReplicas(mount, idx)
- bal.logf("%s: added %d replicas", mount, len(idx))
+ bal.logf("%s: added %d entries to map at %dx (%d replicas)", mount, len(idx), mount.Replication, len(idx)*mount.Replication)
}
bal.logf("mount %s: index done", mounts[0])
}(mounts)
for coll := range collQ {
err := bal.addCollection(coll)
if err != nil {
- errs <- err
+ select {
+ case errs <- err:
+ default:
+ }
for range collQ {
}
return
})
close(collQ)
if err != nil {
- errs <- err
+ select {
+ case errs <- err:
+ default:
+ }
}
}()
func (bal *Balancer) ComputeChangeSets() {
// This just calls balanceBlock() once for each block, using a
// pool of worker goroutines.
- defer timeMe(bal.Logger, "ComputeChangeSets")()
+ defer bal.time("changeset_compute", "wall clock time to compute changesets")()
bal.setupLookupTables()
type balanceTask struct {
slots = append(slots, slot{
mnt: mnt,
repl: repl,
- want: repl != nil && (mnt.ReadOnly || repl.Mtime >= bal.MinMtime),
+ want: repl != nil && mnt.ReadOnly,
})
}
}
have := 0
for _, slot := range slots {
if slot.repl != nil && bal.mountsByClass[class][slot.mnt] && !countedDev[slot.mnt.DeviceID] {
- have++
+ have += slot.mnt.Replication
if slot.mnt.DeviceID != "" {
countedDev[slot.mnt.DeviceID] = true
}
// Prefer a mount that satisfies the
// desired class.
return bal.mountsByClass[class][si.mnt]
- } else if wanti, wantj := si.want, si.want; wanti != wantj {
+ } else if si.want != sj.want {
// Prefer a mount that will have a
// replica no matter what we do here
// -- either because it already has an
// untrashable replica, or because we
// already need it to satisfy a
// different storage class.
- return slots[i].want
+ return si.want
} else if orderi, orderj := srvRendezvous[si.mnt.KeepService], srvRendezvous[sj.mnt.KeepService]; orderi != orderj {
// Prefer a better rendezvous
// position.
// trashing replicas that aren't optimal positions for
// any storage class.
protMnt := map[*KeepMount]bool{}
+ // Replication planned so far (corresponds to wantMnt).
+ replWant := 0
+ // Protected replication (corresponds to protMnt).
+ replProt := 0
// trySlot tries using a slot to meet requirements,
// and returns true if all requirements are met.
trySlot := func(i int) bool {
slot := slots[i]
- if wantDev[slot.mnt.DeviceID] {
+ if wantMnt[slot.mnt] || wantDev[slot.mnt.DeviceID] {
// Already allocated a replica to this
// backend device, possibly on a
// different server.
return false
}
- if len(protMnt) < desired && slot.repl != nil {
+ if replProt < desired && slot.repl != nil && !protMnt[slot.mnt] {
unsafeToDelete[slot.repl.Mtime] = true
protMnt[slot.mnt] = true
+ replProt += slot.mnt.Replication
}
- if len(wantMnt) < desired && (slot.repl != nil || !slot.mnt.ReadOnly) {
+ if replWant < desired && (slot.repl != nil || !slot.mnt.ReadOnly) {
slots[i].want = true
wantSrv[slot.mnt.KeepService] = true
wantMnt[slot.mnt] = true
if slot.mnt.DeviceID != "" {
wantDev[slot.mnt.DeviceID] = true
}
+ replWant += slot.mnt.Replication
}
- return len(protMnt) >= desired && len(wantMnt) >= desired
+ return replProt >= desired && replWant >= desired
}
// First try to achieve desired replication without
if slot.repl == nil || !bal.mountsByClass[class][slot.mnt] {
continue
}
- if safe++; safe >= desired {
+ if safe += slot.mnt.Replication; safe >= desired {
break
}
}
countedDev := map[string]bool{}
var have, want int
for _, slot := range slots {
+ if countedDev[slot.mnt.DeviceID] {
+ continue
+ }
if slot.want {
- want++
+ want += slot.mnt.Replication
}
- if slot.repl != nil && !countedDev[slot.mnt.DeviceID] {
- have++
- if slot.mnt.DeviceID != "" {
- countedDev[slot.mnt.DeviceID] = true
- }
+ if slot.repl != nil {
+ have += slot.mnt.Replication
+ }
+ if slot.mnt.DeviceID != "" {
+ countedDev[slot.mnt.DeviceID] = true
}
}
// TODO: request a Touch if Mtime is duplicated.
var change int
switch {
- case !underreplicated && slot.repl != nil && !slot.want && !unsafeToDelete[slot.repl.Mtime]:
+ case !underreplicated && !slot.want && slot.repl != nil && slot.repl.Mtime < bal.MinMtime && !unsafeToDelete[slot.repl.Mtime]:
slot.mnt.KeepService.AddTrash(Trash{
SizedDigest: blkid,
Mtime: slot.repl.Mtime,
trashes int
replHistogram []int
classStats map[string]replicationStats
+
+ // collectionBytes / collectionBlockBytes = deduplication ratio
+ collectionBytes int64 // sum(bytes in referenced blocks) across all collections
+ collectionBlockBytes int64 // sum(block size) across all blocks referenced by collections
+ collectionBlockRefs int64 // sum(number of blocks referenced) across all collections
+ collectionBlocks int64 // number of blocks referenced by any collection
+}
+
+func (s *balancerStats) dedupByteRatio() float64 {
+ if s.collectionBlockBytes == 0 {
+ return 0
+ }
+ return float64(s.collectionBytes) / float64(s.collectionBlockBytes)
+}
+
+func (s *balancerStats) dedupBlockRatio() float64 {
+ if s.collectionBlocks == 0 {
+ return 0
+ }
+ return float64(s.collectionBlockRefs) / float64(s.collectionBlocks)
}
type replicationStats struct {
surplus := result.have - result.want
bytes := result.blkid.Size()
+ if rc := int64(result.blk.RefCount); rc > 0 {
+ s.collectionBytes += rc * bytes
+ s.collectionBlockBytes += bytes
+ s.collectionBlockRefs += rc
+ s.collectionBlocks++
+ }
+
for class, state := range result.classState {
cs := s.classStats[class]
if state.unachievable {
case surplus > 0:
s.overrep.replicas += surplus
s.overrep.blocks++
- s.overrep.bytes += bytes * int64(len(result.blk.Replicas)-result.want)
+ s.overrep.bytes += bytes * int64(result.have-result.want)
default:
s.justright.replicas += result.want
s.justright.blocks++
s.desired.blocks++
s.desired.bytes += bytes * int64(result.want)
}
- if len(result.blk.Replicas) > 0 {
- s.current.replicas += len(result.blk.Replicas)
+ if result.have > 0 {
+ s.current.replicas += result.have
s.current.blocks++
- s.current.bytes += bytes * int64(len(result.blk.Replicas))
+ s.current.bytes += bytes * int64(result.have)
}
- for len(s.replHistogram) <= len(result.blk.Replicas) {
+ for len(s.replHistogram) <= result.have {
s.replHistogram = append(s.replHistogram, 0)
}
- s.replHistogram[len(result.blk.Replicas)]++
+ s.replHistogram[result.have]++
}
for _, srv := range bal.KeepServices {
s.pulls += len(srv.ChangeSet.Pulls)
s.trashes += len(srv.ChangeSet.Trashes)
}
bal.stats = s
+ bal.Metrics.UpdateStats(s)
}
// PrintStatistics writes statistics about the computed changes to
// existing blocks that are either underreplicated or poorly
// distributed according to rendezvous hashing.
func (bal *Balancer) CommitPulls(c *arvados.Client) error {
+ defer bal.time("send_pull_lists", "wall clock time to send pull lists")()
return bal.commitAsync(c, "send pull list",
func(srv *KeepService) error {
return srv.CommitPulls(c)
// keepstore servers. This has the effect of deleting blocks that are
// overreplicated or unreferenced.
func (bal *Balancer) CommitTrash(c *arvados.Client) error {
+ defer bal.time("send_trash_lists", "wall clock time to send trash lists")()
return bal.commitAsync(c, "send trash list",
func(srv *KeepService) error {
return srv.CommitTrash(c)
var err error
defer func() { errs <- err }()
label := fmt.Sprintf("%s: %v", srv, label)
- defer timeMe(bal.Logger, label)()
err = f(srv)
if err != nil {
err = fmt.Errorf("%s: %v", label, err)
}
}
+func (bal *Balancer) time(name, help string) func() {
+ observer := bal.Metrics.DurationObserver(name+"_seconds", help)
+ t0 := time.Now()
+ bal.Logger.Printf("%s: start", name)
+ return func() {
+ dur := time.Since(t0)
+ observer.Observe(dur.Seconds())
+ bal.Logger.Printf("%s: took %vs", name, dur.Seconds())
+ }
+}
+
// Rendezvous hash sort function. Less efficient than sorting on
// precomputed rendezvous hashes, but also rarely used.
func rendezvousLess(i, j string, blkid arvados.SizedDigest) bool {