From b8461c79fc72d3fc3ed8d0815058e12385eaa4d1 Mon Sep 17 00:00:00 2001 From: Helin Wang Date: Tue, 8 Aug 2017 00:00:06 +0000 Subject: [PATCH] implement init parameters selection with etcd --- go/glide.lock | 10 +- go/master/service_test.go | 4 +- go/pserver/client/c/cclient.go | 20 ++- go/pserver/client/client.go | 10 +- go/pserver/client/client_test.go | 14 +- go/pserver/client/etcd_client.go | 153 ++++++++++++++++--- go/pserver/client/etcd_client_test.go | 97 ++++++++++++ paddle/trainer/NewRemoteParameterUpdater.cpp | 2 +- 8 files changed, 269 insertions(+), 41 deletions(-) create mode 100644 go/pserver/client/etcd_client_test.go diff --git a/go/glide.lock b/go/glide.lock index be1fb24d77..1ecdd21752 100644 --- a/go/glide.lock +++ b/go/glide.lock @@ -1,5 +1,5 @@ hash: 1b9b07408ca7fac27a374dc2ccd2433e4bff090484008a037df967284949a582 -updated: 2017-08-03T21:46:51.744995189Z +updated: 2017-08-07T23:37:48.867469328Z imports: - name: github.com/beorn7/perks version: 4c0e84591b9aa9e6dcfdf3e020114cd81f89d5f9 @@ -10,7 +10,7 @@ imports: - name: github.com/cockroachdb/cmux version: 112f0506e7743d64a6eb8fedbcff13d9979bbf92 - name: github.com/coreos/etcd - version: c31bec0f29facff13f7c3e3d948e55dd6689ed42 + version: d0d1a87aa96ae14914751d42264262cb69eda170 subpackages: - alarm - auth @@ -24,6 +24,7 @@ imports: - error - etcdserver - etcdserver/api + - etcdserver/api/etcdhttp - etcdserver/api/v2http - etcdserver/api/v2http/httptypes - etcdserver/api/v3client @@ -210,11 +211,6 @@ testImports: version: 04cdfd42973bb9c8589fd6a731800cf222fde1a9 subpackages: - spew -- name: github.com/docker/docker - version: b6d164e6c46d8115b146e4c3ac93784e9ef8b49e - subpackages: - - pkg/ioutils - - pkg/longpath - name: github.com/pmezard/go-difflib version: d8ed2627bdf02c080bf22230dbb337003b7aba2d subpackages: diff --git a/go/master/service_test.go b/go/master/service_test.go index 5f91910ecc..87ce9c9043 100644 --- a/go/master/service_test.go +++ b/go/master/service_test.go @@ -1,6 +1,7 @@ package master_test import ( + "io/ioutil" "os" "testing" "time" @@ -8,13 +9,12 @@ import ( "github.com/PaddlePaddle/Paddle/go/master" "github.com/coreos/etcd/clientv3" "github.com/coreos/etcd/embed" - "github.com/docker/docker/pkg/ioutils" "github.com/stretchr/testify/assert" ) func TestNewServiceWithEtcd(t *testing.T) { // setup an embed etcd server - etcdDir, err := ioutils.TempDir("", "") + etcdDir, err := ioutil.TempDir("", "") if err != nil { t.Fatal(err) } diff --git a/go/pserver/client/c/cclient.go b/go/pserver/client/c/cclient.go index 14ad077455..a49cd01522 100644 --- a/go/pserver/client/c/cclient.go +++ b/go/pserver/client/c/cclient.go @@ -90,8 +90,12 @@ func cArrayToSlice(p unsafe.Pointer, len int) []byte { type selector bool -func (s selector) Select() bool { - return bool(s) +func (s selector) Select() (bool, error) { + return bool(s), nil +} + +func (s selector) Done() error { + return nil } type lister []client.Server @@ -114,11 +118,10 @@ func paddle_new_pserver_client(addrs *C.char, selected int) C.paddle_pserver_cli } //export paddle_new_etcd_pserver_client -func paddle_new_etcd_pserver_client(etcdEndpoints *C.char, selected int) C.paddle_pserver_client { - // TODO(Longfei: use etcd lock to decide which trainer to initialize the parameters) +func paddle_new_etcd_pserver_client(etcdEndpoints *C.char) C.paddle_pserver_client { addr := C.GoString(etcdEndpoints) etcdClient := client.NewEtcd(addr) - c := client.NewClient(etcdClient, etcdClient.Desired(), selector(selected != 0)) + c := client.NewClient(etcdClient, etcdClient.Desired(), etcdClient) return add(c) } @@ -136,7 +139,12 @@ func paddle_pserver_client_release(client C.paddle_pserver_client) { //export paddle_begin_init_params func paddle_begin_init_params(client C.paddle_pserver_client) C.int { c := get(client) - if selected := c.BeginInitParams(); selected { + selected, err := c.BeginInitParams() + if err != nil { + panic(err) + } + + if selected { return 1 } return 0 diff --git a/go/pserver/client/client.go b/go/pserver/client/client.go index 15adda4735..20d91e7703 100644 --- a/go/pserver/client/client.go +++ b/go/pserver/client/client.go @@ -27,9 +27,13 @@ import ( // TODO(helin): add RPC call retry logic -// Selector selects if the client should initialize parameter servers. +// Selector selects if the client should initialize parameters and +// reports the initialization process done. type Selector interface { - Select() bool + // Select selects if the client should initialize parameter servers. + Select() (bool, error) + // Done indicates the initialization process is done. + Done() error } // Server is the identification of a parameter Server. @@ -115,7 +119,7 @@ func (c *Client) monitorPservers(l Lister, pserverNum int) { // servers. Other trainers will be blocked until the initialization is // done, and they need to get the initialized parameters from // parameter servers using GetParams. -func (c *Client) BeginInitParams() bool { +func (c *Client) BeginInitParams() (bool, error) { return c.sel.Select() } diff --git a/go/pserver/client/client_test.go b/go/pserver/client/client_test.go index 1243ebd683..c3d88e926d 100644 --- a/go/pserver/client/client_test.go +++ b/go/pserver/client/client_test.go @@ -124,8 +124,12 @@ func initEtcdClient() { type selector bool -func (s selector) Select() bool { - return bool(s) +func (s selector) Select() (bool, error) { + return bool(s), nil +} + +func (s selector) Done() error { + return nil } type lister []client.Server @@ -135,7 +139,11 @@ func (l lister) List() []client.Server { } func testClient(t *testing.T, c *client.Client) { - selected := c.BeginInitParams() + selected, err := c.BeginInitParams() + if err != nil { + t.Fatal(err) + } + if !selected { t.Fatal("should be selected.") } diff --git a/go/pserver/client/etcd_client.go b/go/pserver/client/etcd_client.go index 977ae5af37..f9071caaa8 100644 --- a/go/pserver/client/etcd_client.go +++ b/go/pserver/client/etcd_client.go @@ -16,53 +16,60 @@ package client import ( "context" + "errors" + "fmt" "strconv" "strings" "time" "github.com/PaddlePaddle/Paddle/go/pserver" "github.com/coreos/etcd/clientv3" + "github.com/coreos/etcd/clientv3/concurrency" log "github.com/sirupsen/logrus" ) const ( defaultEtcdTimeout time.Duration = 5 * time.Second + + initLockPath = "/init_ps/lock" + initDonePath = "/init_ps/done" + initDoneVal = "1" ) -// EtcdClient is used by pserver client that is a part of trainer process. +// Etcd is used by pserver client that is a part of trainer process. // TODO: -// 1. add watcher to watch the change state of pservers) -// 1. add etcd lock) -type EtcdClient struct { +// 1. add watcher to watch the change state of pservers. +type Etcd struct { client *clientv3.Client timeout time.Duration endpoints []string + lock *concurrency.Mutex } // Desired read ps desired number from etcd. -func (p *EtcdClient) Desired() int { +func (e *Etcd) Desired() int { var psDesired int for { - ctx, cancel := context.WithTimeout(context.Background(), p.timeout) - resp, err := p.client.Get(ctx, pserver.PsDesired) + ctx, cancel := context.WithTimeout(context.Background(), e.timeout) + resp, err := e.client.Get(ctx, pserver.PsDesired) cancel() if err != nil { log.Errorf("Get ps dresire number failed! recnnectiong..., %v", err) - time.Sleep(p.timeout) + time.Sleep(e.timeout) continue } kvs := resp.Kvs if len(kvs) == 0 { log.Infoln("Waiting for ps desired registered ...") - time.Sleep(p.timeout) + time.Sleep(e.timeout) continue } psDesired, err = strconv.Atoi(string(resp.Kvs[0].Value)) if err != nil { log.Errorf("psDesired %d invalid %v", psDesired, err) - time.Sleep(p.timeout) + time.Sleep(e.timeout) continue } @@ -73,26 +80,26 @@ func (p *EtcdClient) Desired() int { } // List return the pserver list read from etcd. -func (p *EtcdClient) List() []Server { - psDesired := p.Desired() +func (e *Etcd) List() []Server { + psDesired := e.Desired() servers := make([]Server, psDesired) for { for i := 0; i < psDesired; i++ { - ctx, cancel := context.WithTimeout(context.Background(), p.timeout) + ctx, cancel := context.WithTimeout(context.Background(), e.timeout) psKey := pserver.PsPath + strconv.Itoa(i) log.Debugf("checking %s", psKey) - resp, err := p.client.Get(ctx, psKey) + resp, err := e.client.Get(ctx, psKey) cancel() if err != nil { log.Infof("Get psKey= %s error, %v", psKey, err) - time.Sleep(p.timeout) + time.Sleep(e.timeout) continue } kvs := resp.Kvs if len(kvs) == 0 { log.Infof("Waiting for ps addr registered ...") - time.Sleep(p.timeout) + time.Sleep(e.timeout) continue } @@ -100,7 +107,7 @@ func (p *EtcdClient) List() []Server { // TODO(Longfei) check the ps address if psAddr == "" { log.Infof("Get psKey = %s, psAddr is empty", psKey) - time.Sleep(p.timeout) + time.Sleep(e.timeout) continue } log.Debugf("got value (%s) for key: %s", psAddr, psKey) @@ -113,7 +120,7 @@ func (p *EtcdClient) List() []Server { } // NewEtcd create a etcd client to return the state of pserver on etcd. -func NewEtcd(endpoints string) *EtcdClient { +func NewEtcd(endpoints string) *Etcd { ep := strings.Split(endpoints, ",") var cli *clientv3.Client var err error @@ -130,10 +137,118 @@ func NewEtcd(endpoints string) *EtcdClient { break } log.Infof("Connected to etcd: %s\n", endpoints) - client := &EtcdClient{ + client := &Etcd{ client: cli, timeout: defaultEtcdTimeout, endpoints: ep, } return client } + +// Select indicates if the current trainer is selected to initialize +// the pserver parameters. +func (e *Etcd) Select() (bool, error) { + sess, err := concurrency.NewSession(e.client, concurrency.WithTTL(5)) + if err != nil { + return false, err + } + + lock := concurrency.NewMutex(sess, initLockPath) + log.Infof("Trying to acquire lock at %s.", initLockPath) + // Do not use timeout context here, since we don't know how + // long does it take for other trainers to initialize the + // parameters. + err = lock.Lock(context.Background()) + if err != nil { + return false, err + } + log.Infof("Successfully acquired lock at %s.", initLockPath) + + get := clientv3.OpGet(initDonePath) + ctx, cancel := context.WithTimeout(context.Background(), e.timeout) + tresp, err := e.client.Txn(ctx).If(lock.IsOwner()).Then(get).Commit() + cancel() + if err != nil { + return false, err + } + + if !tresp.Succeeded { + return false, errors.New("no longer the owner of the lock") + } + + resp := tresp.Responses[0].GetResponseRange() + + if len(resp.Kvs) == 0 { + // Key value not set, select current trainer. + e.lock = lock + log.Infoln("Trainer selected.") + return true, nil + } + + if string(resp.Kvs[0].Value) == initDoneVal { + log.Infoln("Initialization is already done.") + ctx, cancel = context.WithTimeout(context.Background(), e.timeout) + err = lock.Unlock(ctx) + cancel() + if err != nil { + log.Errorln(err) + } + return false, nil + } + + return false, fmt.Errorf("key %s have unexpected value: %v", initDonePath, resp.Kvs[0].Value) +} + +// Done indicates the parameter initialization process is done. +func (e *Etcd) Done() error { + if e.lock == nil { + return errors.New("lock is nil, Done called unexpectedly") + } + + put := clientv3.OpPut(initDonePath, initDoneVal) + ctx, cancel := context.WithTimeout(context.Background(), e.timeout) + tresp, err := e.client.Txn(ctx).If(e.lock.IsOwner()).Then(put).Commit() + cancel() + if err != nil { + return err + } + + if !tresp.Succeeded { + return errors.New("no longer the owner of the lock") + } + + ctx, cancel = context.WithTimeout(context.Background(), e.timeout) + err = e.lock.Unlock(ctx) + cancel() + if err != nil { + log.Errorln(err) + } else { + e.lock = nil + } + + return nil +} + +// Close closes the etcd client. +func (e *Etcd) Close() error { + var err error + if e.lock != nil { + ctx, cancel := context.WithTimeout(context.Background(), e.timeout) + err = e.lock.Unlock(ctx) + cancel() + if err == nil { + e.lock = nil + } + } + + cErr := e.client.Close() + if cErr != nil { + if err != nil { + log.Errorln(cErr) + return err + } + return cErr + } + + return err +} diff --git a/go/pserver/client/etcd_client_test.go b/go/pserver/client/etcd_client_test.go new file mode 100644 index 0000000000..9daeb97d4c --- /dev/null +++ b/go/pserver/client/etcd_client_test.go @@ -0,0 +1,97 @@ +package client_test + +import ( + "io/ioutil" + "os" + "sync" + "testing" + + "github.com/PaddlePaddle/Paddle/go/pserver/client" + "github.com/coreos/etcd/embed" +) + +func TestSelector(t *testing.T) { + etcdDir, err := ioutil.TempDir("", "") + if err != nil { + t.Fatal(err) + } + cfg := embed.NewConfig() + cfg.Dir = etcdDir + e, err := embed.StartEtcd(cfg) + if err != nil { + t.Fatal(err) + } + + defer func() { + e.Close() + if err := os.RemoveAll(etcdDir); err != nil { + t.Fatal(err) + } + }() + + <-e.Server.ReadyNotify() + + var mu sync.Mutex + selectedCount := 0 + var wg sync.WaitGroup + selectAndDone := func(c *client.Etcd) { + defer wg.Done() + + selected, err := c.Select() + if err != nil { + panic(err) + } + + if selected { + mu.Lock() + selectedCount++ + mu.Unlock() + err = c.Done() + if err != nil { + t.Fatal(err) + } + } + } + + c0 := client.NewEtcd("127.0.0.1:2379") + c1 := client.NewEtcd("127.0.0.1:2379") + c2 := client.NewEtcd("127.0.0.1:2379") + c3 := client.NewEtcd("127.0.0.1:2379") + wg.Add(3) + go selectAndDone(c0) + go selectAndDone(c1) + go selectAndDone(c2) + wg.Wait() + + // simulate trainer crashed and restarted after the + // initialization process. + wg.Add(1) + go selectAndDone(c3) + wg.Wait() + + mu.Lock() + if selectedCount != 1 { + t.Fatal("selected count wrong:", selectedCount) + } + mu.Unlock() + + err = c0.Close() + if err != nil { + t.Fatal(err) + } + + err = c1.Close() + if err != nil { + t.Fatal(err) + } + + err = c2.Close() + if err != nil { + t.Fatal(err) + } + + err = c3.Close() + if err != nil { + t.Fatal(err) + } +} diff --git a/paddle/trainer/NewRemoteParameterUpdater.cpp b/paddle/trainer/NewRemoteParameterUpdater.cpp index e1558e3fdf..31bbab3aef 100644 --- a/paddle/trainer/NewRemoteParameterUpdater.cpp +++ b/paddle/trainer/NewRemoteParameterUpdater.cpp @@ -51,7 +51,7 @@ void NewRemoteParameterUpdater::init( // create parameter server client. if (useEtcd_) { parameterClient_ = paddle_new_etcd_pserver_client( - (char *)pserverSpec_.c_str(), FLAGS_trainer_id == 0); + (char *)pserverSpec_.c_str()); } else { parameterClient_ = paddle_new_pserver_client((char *)pserverSpec_.c_str(), FLAGS_trainer_id == 0); -- GitLab