services/crunch-dispatch-local/crunch-dispatch-local.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package main
   6
   7 // Dispatcher service for Crunch that runs containers locally.
   8
   9 import (
  10         "context"
  11         "flag"
  12         "fmt"
  13         "os"
  14         "os/exec"
  15         "os/signal"
  16         "runtime"
  17         "sync"
  18         "syscall"
  19         "time"
  20
  21         "git.arvados.org/arvados.git/lib/cmd"
  22         "git.arvados.org/arvados.git/lib/config"
  23         "git.arvados.org/arvados.git/sdk/go/arvados"
  24         "git.arvados.org/arvados.git/sdk/go/arvadosclient"
  25         "git.arvados.org/arvados.git/sdk/go/dispatch"
  26         "github.com/pbnjay/memory"
  27         "github.com/sirupsen/logrus"
  28 )
  29
  30 var version = "dev"
  31
  32 var (
  33         runningCmds      map[string]*exec.Cmd
  34         runningCmdsMutex sync.Mutex
  35         waitGroup        sync.WaitGroup
  36         crunchRunCommand string
  37 )
  38
  39 func main() {
  40         baseLogger := logrus.StandardLogger()
  41         if os.Getenv("DEBUG") != "" {
  42                 baseLogger.SetLevel(logrus.DebugLevel)
  43         }
  44         baseLogger.Formatter = &logrus.JSONFormatter{
  45                 TimestampFormat: "2006-01-02T15:04:05.000000000Z07:00",
  46         }
  47
  48         flags := flag.NewFlagSet("crunch-dispatch-local", flag.ExitOnError)
  49
  50         pollInterval := flags.Int(
  51                 "poll-interval",
  52                 10,
  53                 "Interval in seconds to poll for queued containers")
  54
  55         flags.StringVar(&crunchRunCommand,
  56                 "crunch-run-command",
  57                 "",
  58                 "Crunch command to run container")
  59
  60         getVersion := flags.Bool(
  61                 "version",
  62                 false,
  63                 "Print version information and exit.")
  64
  65         if ok, code := cmd.ParseFlags(flags, os.Args[0], os.Args[1:], "", os.Stderr); !ok {
  66                 os.Exit(code)
  67         }
  68
  69         // Print version information if requested
  70         if *getVersion {
  71                 fmt.Printf("crunch-dispatch-local %s\n", version)
  72                 return
  73         }
  74
  75         loader := config.NewLoader(nil, baseLogger)
  76         cfg, err := loader.Load()
  77         if err != nil {
  78                 fmt.Fprintf(os.Stderr, "error loading config: %s\n", err)
  79                 os.Exit(1)
  80         }
  81         cluster, err := cfg.GetCluster("")
  82         if err != nil {
  83                 fmt.Fprintf(os.Stderr, "config error: %s\n", err)
  84                 os.Exit(1)
  85         }
  86
  87         if crunchRunCommand == "" {
  88                 crunchRunCommand = cluster.Containers.CrunchRunCommand
  89         }
  90
  91         logger := baseLogger.WithField("ClusterID", cluster.ClusterID)
  92         logger.Printf("crunch-dispatch-local %s started", version)
  93
  94         runningCmds = make(map[string]*exec.Cmd)
  95
  96         var client arvados.Client
  97         client.APIHost = cluster.Services.Controller.ExternalURL.Host
  98         client.AuthToken = cluster.SystemRootToken
  99         client.Insecure = cluster.TLS.Insecure
 100
 101         if client.APIHost != "" || client.AuthToken != "" {
 102                 // Copy real configs into env vars so [a]
 103                 // MakeArvadosClient() uses them, and [b] they get
 104                 // propagated to crunch-run via SLURM.
 105                 os.Setenv("ARVADOS_API_HOST", client.APIHost)
 106                 os.Setenv("ARVADOS_API_TOKEN", client.AuthToken)
 107                 os.Setenv("ARVADOS_API_HOST_INSECURE", "")
 108                 if client.Insecure {
 109                         os.Setenv("ARVADOS_API_HOST_INSECURE", "1")
 110                 }
 111         } else {
 112                 logger.Warnf("Client credentials missing from config, so falling back on environment variables (deprecated).")
 113         }
 114
 115         arv, err := arvadosclient.MakeArvadosClient()
 116         if err != nil {
 117                 logger.Errorf("error making Arvados client: %v", err)
 118                 os.Exit(1)
 119         }
 120         arv.Retries = 25
 121
 122         ctx, cancel := context.WithCancel(context.Background())
 123
 124         localRun := LocalRun{startFunc, make(chan ResourceRequest), ctx, cluster}
 125
 126         go localRun.throttle(logger)
 127
 128         dispatcher := dispatch.Dispatcher{
 129                 Logger:       logger,
 130                 Arv:          arv,
 131                 RunContainer: localRun.run,
 132                 PollPeriod:   time.Duration(*pollInterval) * time.Second,
 133         }
 134
 135         err = dispatcher.Run(ctx)
 136         if err != nil {
 137                 logger.Error(err)
 138                 return
 139         }
 140
 141         c := make(chan os.Signal, 1)
 142         signal.Notify(c, os.Interrupt, syscall.SIGTERM, syscall.SIGQUIT)
 143         sig := <-c
 144         logger.Printf("Received %s, shutting down", sig)
 145         signal.Stop(c)
 146
 147         cancel()
 148
 149         runningCmdsMutex.Lock()
 150         // Finished dispatching; interrupt any crunch jobs that are still running
 151         for _, cmd := range runningCmds {
 152                 cmd.Process.Signal(os.Interrupt)
 153         }
 154         runningCmdsMutex.Unlock()
 155
 156         // Wait for all running crunch jobs to complete / terminate
 157         waitGroup.Wait()
 158 }
 159
 160 func startFunc(container arvados.Container, cmd *exec.Cmd) error {
 161         return cmd.Start()
 162 }
 163
 164 type ResourceRequest struct {
 165         uuid  string
 166         vcpus int
 167         ram   int64
 168         gpus  int
 169         ready chan bool
 170 }
 171
 172 type LocalRun struct {
 173         startCmd         func(container arvados.Container, cmd *exec.Cmd) error
 174         concurrencyLimit chan ResourceRequest
 175         ctx              context.Context
 176         cluster          *arvados.Cluster
 177 }
 178
 179 func (lr *LocalRun) throttle(logger logrus.FieldLogger) {
 180         maxVcpus := runtime.NumCPU()
 181         var maxRam int64 = int64(memory.TotalMemory())
 182
 183         // treat all GPUs as a single resource for now.
 184         maxGpus := 1
 185
 186         var allocVcpus int
 187         var allocRam int64
 188         var allocGpus int
 189
 190         pending := []ResourceRequest{}
 191
 192 NextEvent:
 193         for {
 194                 rr := <-lr.concurrencyLimit
 195
 196                 if rr.vcpus > 0 {
 197                         // allocating resources
 198                         pending = append(pending, rr)
 199                 } else {
 200                         // releasing resources (these should be
 201                         // negative numbers)
 202                         allocVcpus += rr.vcpus
 203                         allocRam += rr.ram
 204                         allocGpus += rr.gpus
 205
 206                         logger.Infof("%v removed allocation (cpus: %v ram: %v gpus: %v); total allocated (cpus: %v ram: %v gpus: %v)",
 207                                 rr.uuid, rr.vcpus, rr.ram, rr.gpus,
 208                                 allocVcpus, allocRam, allocGpus)
 209                 }
 210
 211                 for len(pending) > 0 {
 212                         rr := pending[0]
 213                         if rr.vcpus > maxVcpus || rr.ram > maxRam || rr.gpus > maxGpus {
 214                                 // resource request can never be fulfilled
 215                                 rr.ready <- false
 216                                 continue
 217                         }
 218
 219                         if (allocVcpus+rr.vcpus) > maxVcpus || (allocRam+rr.ram) > maxRam || (allocGpus+rr.gpus) > maxGpus {
 220                                 logger.Info("Insufficient resources to start %v, waiting for next event", rr.uuid)
 221                                 // can't be scheduled yet, go up to
 222                                 // the top and wait for the next event
 223                                 continue NextEvent
 224                         }
 225
 226                         allocVcpus += rr.vcpus
 227                         allocRam += rr.ram
 228                         allocGpus += rr.gpus
 229                         rr.ready <- true
 230
 231                         logger.Infof("%v added allocation (cpus: %v ram: %v gpus: %v); total allocated (cpus: %v ram: %v gpus: %v)",
 232                                 rr.uuid, rr.vcpus, rr.ram, rr.gpus,
 233                                 allocVcpus, allocRam, allocGpus)
 234
 235                         // shift up
 236                         for i := 0; i < len(pending)-1; i++ {
 237                                 pending[i] = pending[i+1]
 238                         }
 239                         pending = pending[0 : len(pending)-1]
 240                 }
 241
 242         }
 243 }
 244
 245 // Run a container.
 246 //
 247 // If the container is Locked, start a new crunch-run process and wait until
 248 // crunch-run completes.  If the priority is set to zero, set an interrupt
 249 // signal to the crunch-run process.
 250 //
 251 // If the container is in any other state, or is not Complete/Cancelled after
 252 // crunch-run terminates, mark the container as Cancelled.
 253 func (lr *LocalRun) run(dispatcher *dispatch.Dispatcher,
 254         container arvados.Container,
 255         status <-chan arvados.Container) error {
 256
 257         uuid := container.UUID
 258
 259         if container.State == dispatch.Locked {
 260
 261                 resourceRequest := ResourceRequest{
 262                         uuid:  container.UUID,
 263                         vcpus: container.RuntimeConstraints.VCPUs,
 264                         ram: (container.RuntimeConstraints.RAM +
 265                                 container.RuntimeConstraints.KeepCacheRAM +
 266                                 int64(lr.cluster.Containers.ReserveExtraRAM)),
 267                         gpus:  container.RuntimeConstraints.CUDA.DeviceCount,
 268                         ready: make(chan bool)}
 269
 270                 select {
 271                 case lr.concurrencyLimit <- resourceRequest:
 272                         break
 273                 case <-lr.ctx.Done():
 274                         return lr.ctx.Err()
 275                 }
 276
 277                 canRun := <-resourceRequest.ready
 278
 279                 if !canRun {
 280                         dispatcher.Logger.Warnf("Container resource request %v cannot be fulfilled.", uuid)
 281                         dispatcher.UpdateState(uuid, dispatch.Cancelled)
 282                         return nil
 283                 }
 284
 285                 defer func() {
 286                         resourceRequest.vcpus = -resourceRequest.vcpus
 287                         resourceRequest.ram = -resourceRequest.ram
 288                         resourceRequest.gpus = -resourceRequest.gpus
 289                         lr.concurrencyLimit <- resourceRequest
 290                 }()
 291
 292                 select {
 293                 case c := <-status:
 294                         // Check for state updates after possibly
 295                         // waiting to be ready-to-run
 296                         if c.Priority == 0 {
 297                                 goto Finish
 298                         }
 299                 default:
 300                         break
 301                 }
 302
 303                 waitGroup.Add(1)
 304                 defer waitGroup.Done()
 305
 306                 args := []string{"--runtime-engine=" + lr.cluster.Containers.RuntimeEngine}
 307                 args = append(args, lr.cluster.Containers.CrunchRunArgumentsList...)
 308                 args = append(args, uuid)
 309
 310                 cmd := exec.Command(crunchRunCommand, args...)
 311                 cmd.Stdin = nil
 312                 cmd.Stderr = os.Stderr
 313                 cmd.Stdout = os.Stderr
 314
 315                 dispatcher.Logger.Printf("starting container %v", uuid)
 316
 317                 // Add this crunch job to the list of runningCmds only if we
 318                 // succeed in starting crunch-run.
 319
 320                 runningCmdsMutex.Lock()
 321                 if err := lr.startCmd(container, cmd); err != nil {
 322                         runningCmdsMutex.Unlock()
 323                         dispatcher.Logger.Warnf("error starting %q for %s: %s", crunchRunCommand, uuid, err)
 324                         dispatcher.UpdateState(uuid, dispatch.Cancelled)
 325                 } else {
 326                         runningCmds[uuid] = cmd
 327                         runningCmdsMutex.Unlock()
 328
 329                         // Need to wait for crunch-run to exit
 330                         done := make(chan struct{})
 331
 332                         go func() {
 333                                 if _, err := cmd.Process.Wait(); err != nil {
 334                                         dispatcher.Logger.Warnf("error while waiting for crunch job to finish for %v: %q", uuid, err)
 335                                 }
 336                                 dispatcher.Logger.Debugf("sending done")
 337                                 done <- struct{}{}
 338                         }()
 339
 340                 Loop:
 341                         for {
 342                                 select {
 343                                 case <-done:
 344                                         break Loop
 345                                 case c := <-status:
 346                                         // Interrupt the child process if priority changes to 0
 347                                         if (c.State == dispatch.Locked || c.State == dispatch.Running) && c.Priority == 0 {
 348                                                 dispatcher.Logger.Printf("sending SIGINT to pid %d to cancel container %v", cmd.Process.Pid, uuid)
 349                                                 cmd.Process.Signal(os.Interrupt)
 350                                         }
 351                                 }
 352                         }
 353                         close(done)
 354
 355                         dispatcher.Logger.Printf("finished container run for %v", uuid)
 356
 357                         // Remove the crunch job from runningCmds
 358                         runningCmdsMutex.Lock()
 359                         delete(runningCmds, uuid)
 360                         runningCmdsMutex.Unlock()
 361                 }
 362         }
 363
 364 Finish:
 365
 366         // If the container is not finalized, then change it to "Cancelled".
 367         err := dispatcher.Arv.Get("containers", uuid, nil, &container)
 368         if err != nil {
 369                 dispatcher.Logger.Warnf("error getting final container state: %v", err)
 370         }
 371         if container.State == dispatch.Locked || container.State == dispatch.Running {
 372                 dispatcher.Logger.Warnf("after %q process termination, container state for %v is %q; updating it to %q",
 373                         crunchRunCommand, uuid, container.State, dispatch.Cancelled)
 374                 dispatcher.UpdateState(uuid, dispatch.Cancelled)
 375         }
 376
 377         // drain any subsequent status changes
 378         for range status {
 379         }
 380
 381         dispatcher.Logger.Printf("finalized container %v", uuid)
 382         return nil
 383 }