X-Git-Url: https://git.arvados.org/lightning.git/blobdiff_plain/632ce6b4c5bc7d0b5fb08c71e361b996bd177464..08aa5f2c415150de2738378b6d4fbce88a9e6a59:/arvados.go diff --git a/arvados.go b/arvados.go index f2819d01ef..68f275508d 100644 --- a/arvados.go +++ b/arvados.go @@ -1,6 +1,11 @@ -package main +// Copyright (C) The Lightning Authors. All rights reserved. +// +// SPDX-License-Identifier: AGPL-3.0 + +package lightning import ( + "bufio" "context" "encoding/json" "errors" @@ -10,6 +15,7 @@ import ( "net/url" "os" "regexp" + "strconv" "strings" "sync" "time" @@ -18,6 +24,7 @@ import ( "git.arvados.org/arvados.git/sdk/go/arvados" "git.arvados.org/arvados.git/sdk/go/arvadosclient" "git.arvados.org/arvados.git/sdk/go/keepclient" + "github.com/klauspost/pgzip" log "github.com/sirupsen/logrus" "golang.org/x/crypto/blake2b" "golang.org/x/net/websocket" @@ -171,7 +178,7 @@ reconnect: } client.mtx.Lock() for ch := range client.notifying[msg.ObjectUUID] { - ch <- msg + go func() { ch <- msg }() } client.mtx.Unlock() } @@ -238,7 +245,7 @@ func (runner *arvadosContainerRunner) RunContext(ctx context.Context) (string, e keepCache = 2 } rc := arvados.RuntimeConstraints{ - API: &runner.APIAccess, + API: runner.APIAccess, VCPUs: runner.VCPUs, RAM: runner.RAM, KeepCacheRAM: (1 << 26) * int64(keepCache) * int64(runner.VCPUs), @@ -261,6 +268,13 @@ func (runner *arvadosContainerRunner) RunContext(ctx context.Context) (string, e "runtime_constraints": rc, "priority": runner.Priority, "state": arvados.ContainerRequestStateCommitted, + "scheduling_parameters": arvados.SchedulingParameters{ + Preemptible: false, + Partitions: []string{}, + }, + "environment": map[string]string{ + "GOMAXPROCS": fmt.Sprintf("%d", rc.VCPUs), + }, }, }) if err != nil { @@ -275,6 +289,7 @@ func (runner *arvadosContainerRunner) RunContext(ctx context.Context) (string, e subscribedUUID := "" defer func() { if subscribedUUID != "" { + log.Printf("unsubscribe container UUID: %s", subscribedUUID) client.Unsubscribe(logch, subscribedUUID) } }() @@ -283,14 +298,18 @@ func (runner *arvadosContainerRunner) RunContext(ctx context.Context) (string, e lastState := cr.State refreshCR := func() { - err = runner.Client.RequestAndDecode(&cr, "GET", "arvados/v1/container_requests/"+cr.UUID, nil, nil) + ctx, cancel := context.WithDeadline(ctx, time.Now().Add(time.Minute)) + defer cancel() + err = runner.Client.RequestAndDecodeContext(ctx, &cr, "GET", "arvados/v1/container_requests/"+cr.UUID, nil, nil) if err != nil { fmt.Fprint(os.Stderr, neednewline) + neednewline = "" log.Printf("error getting container request: %s", err) return } if lastState != cr.State { fmt.Fprint(os.Stderr, neednewline) + neednewline = "" log.Printf("container request state: %s", cr.State) lastState = cr.State } @@ -298,14 +317,16 @@ func (runner *arvadosContainerRunner) RunContext(ctx context.Context) (string, e fmt.Fprint(os.Stderr, neednewline) neednewline = "" if subscribedUUID != "" { + log.Printf("unsubscribe container UUID: %s", subscribedUUID) client.Unsubscribe(logch, subscribedUUID) } + log.Printf("subscribe container UUID: %s", cr.ContainerUUID) client.Subscribe(logch, cr.ContainerUUID) subscribedUUID = cr.ContainerUUID } } - var reCrunchstat = regexp.MustCompile(`mem .* rss`) + var reCrunchstat = regexp.MustCompile(`mem .* (\d+) rss`) waitctr: for cr.State != arvados.ContainerRequestStateFinal { select { @@ -335,9 +356,10 @@ waitctr: } case "crunchstat": for _, line := range strings.Split(msg.Properties.Text, "\n") { - mem := reCrunchstat.FindString(line) - if mem != "" { - fmt.Fprintf(os.Stderr, "%s \r", mem) + m := reCrunchstat.FindStringSubmatch(line) + if m != nil { + rss, _ := strconv.ParseInt(m[1], 10, 64) + fmt.Fprintf(os.Stderr, "%s rss %.3f GB \r", cr.UUID, float64(rss)/1e9) neednewline = "\n" } } @@ -376,16 +398,20 @@ func (runner *arvadosContainerRunner) TranslatePaths(paths ...*string) error { if m == nil { return fmt.Errorf("cannot find uuid in path: %q", *path) } - uuid := m[2] - mnt, ok := runner.Mounts["/mnt/"+uuid] + collID := m[2] + mnt, ok := runner.Mounts["/mnt/"+collID] if !ok { mnt = map[string]interface{}{ "kind": "collection", - "uuid": uuid, } - runner.Mounts["/mnt/"+uuid] = mnt + if len(collID) == 27 { + mnt["uuid"] = collID + } else { + mnt["portable_data_hash"] = collID + } + runner.Mounts["/mnt/"+collID] = mnt } - *path = "/mnt/" + uuid + m[3] + *path = "/mnt/" + collID + m[3] } return nil } @@ -463,9 +489,52 @@ func (runner *arvadosContainerRunner) makeCommandCollection() (string, error) { return coll.UUID, nil } -var arvadosClientFromEnv = arvados.NewClientFromEnv() +// zopen returns a reader for the given file, using the arvados API +// instead of arv-mount/fuse where applicable, and transparently +// decompressing the input if fnm ends with ".gz". +func zopen(fnm string) (io.ReadCloser, error) { + f, err := open(fnm) + if err != nil || !strings.HasSuffix(fnm, ".gz") { + return f, err + } + rdr, err := pgzip.NewReader(bufio.NewReaderSize(f, 4*1024*1024)) + if err != nil { + f.Close() + return nil, err + } + return gzipr{rdr, f}, nil +} + +// gzipr wraps a ReadCloser and a Closer, presenting a single Close() +// method that closes both wrapped objects. +type gzipr struct { + io.ReadCloser + io.Closer +} + +func (gr gzipr) Close() error { + e1 := gr.ReadCloser.Close() + e2 := gr.Closer.Close() + if e1 != nil { + return e1 + } + return e2 +} + +var ( + arvadosClientFromEnv = arvados.NewClientFromEnv() + keepClient *keepclient.KeepClient + siteFS arvados.CustomFileSystem + siteFSMtx sync.Mutex +) -func open(fnm string) (io.ReadCloser, error) { +type file interface { + io.ReadCloser + io.Seeker + Readdir(n int) ([]os.FileInfo, error) +} + +func open(fnm string) (file, error) { if os.Getenv("ARVADOS_API_HOST") == "" { return os.Open(fnm) } @@ -473,32 +542,41 @@ func open(fnm string) (io.ReadCloser, error) { if m == nil { return os.Open(fnm) } - uuid := m[2] - mnt := "/mnt/" + uuid + "/" - if !strings.HasPrefix(fnm, mnt) { - return os.Open(fnm) - } + collectionUUID := m[2] + collectionPath := m[3] - log.Infof("reading %q from %s using Arvados client library", fnm[len(mnt):], uuid) - ac, err := arvadosclient.New(arvadosClientFromEnv) + siteFSMtx.Lock() + defer siteFSMtx.Unlock() + if siteFS == nil { + log.Info("setting up Arvados client") + ac, err := arvadosclient.New(arvadosClientFromEnv) + if err != nil { + return nil, err + } + ac.Client = arvados.DefaultSecureClient + keepClient = keepclient.New(ac) + // Don't use keepclient's default short timeouts. + keepClient.HTTPClient = arvados.DefaultSecureClient + keepClient.BlockCache = &keepclient.BlockCache{MaxBlocks: 4} + siteFS = arvadosClientFromEnv.SiteFileSystem(keepClient) + } else { + keepClient.BlockCache.MaxBlocks += 2 + } + + log.Infof("reading %q from %s using Arvados client", collectionPath, collectionUUID) + f, err := siteFS.Open("by_id/" + collectionUUID + collectionPath) if err != nil { return nil, err } - ac.Client = arvados.DefaultSecureClient - kc := keepclient.New(ac) - // Don't use keepclient's default short timeouts. - kc.HTTPClient = arvados.DefaultSecureClient - // Don't cache more than one block for this file. - kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 1} + return &reduceCacheOnClose{file: f}, nil +} - var coll arvados.Collection - err = arvadosClientFromEnv.RequestAndDecode(&coll, "GET", "arvados/v1/collections/"+uuid, nil, arvados.GetOptions{Select: []string{"uuid", "manifest_text"}}) - if err != nil { - return nil, err - } - fs, err := coll.FileSystem(arvadosClientFromEnv, kc) - if err != nil { - return nil, err - } - return fs.Open(fnm[len(mnt):]) +type reduceCacheOnClose struct { + file + once sync.Once +} + +func (rc *reduceCacheOnClose) Close() error { + rc.once.Do(func() { keepClient.BlockCache.MaxBlocks -= 2 }) + return rc.file.Close() }