services/crunch-dispatch-slurm/crunch-dispatch-slurm_test.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package main
   6
   7 import (
   8         "bytes"
   9         "context"
  10         "errors"
  11         "fmt"
  12         "io"
  13         "io/ioutil"
  14         "net/http"
  15         "net/http/httptest"
  16         "os"
  17         "os/exec"
  18         "strings"
  19         "testing"
  20         "time"
  21
  22         "git.curoverse.com/arvados.git/lib/dispatchcloud"
  23         "git.curoverse.com/arvados.git/sdk/go/arvados"
  24         "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
  25         "git.curoverse.com/arvados.git/sdk/go/arvadostest"
  26         "git.curoverse.com/arvados.git/sdk/go/dispatch"
  27         "github.com/sirupsen/logrus"
  28         . "gopkg.in/check.v1"
  29 )
  30
  31 // Gocheck boilerplate
  32 func Test(t *testing.T) {
  33         TestingT(t)
  34 }
  35
  36 var _ = Suite(&IntegrationSuite{})
  37 var _ = Suite(&StubbedSuite{})
  38
  39 type IntegrationSuite struct {
  40         disp  Dispatcher
  41         slurm slurmFake
  42 }
  43
  44 func (s *IntegrationSuite) SetUpTest(c *C) {
  45         arvadostest.StartAPI()
  46         os.Setenv("ARVADOS_API_TOKEN", arvadostest.Dispatch1Token)
  47         s.disp = Dispatcher{}
  48         s.disp.cluster = &arvados.Cluster{}
  49         s.disp.setup()
  50         s.slurm = slurmFake{}
  51 }
  52
  53 func (s *IntegrationSuite) TearDownTest(c *C) {
  54         arvadostest.ResetEnv()
  55         arvadostest.StopAPI()
  56 }
  57
  58 type slurmFake struct {
  59         didBatch      [][]string
  60         didCancel     []string
  61         didRelease    []string
  62         didRenice     [][]string
  63         queue         string
  64         rejectNice10K bool
  65         // If non-nil, run this func during the 2nd+ call to Cancel()
  66         onCancel func()
  67         // Error returned by Batch()
  68         errBatch error
  69 }
  70
  71 func (sf *slurmFake) Batch(script io.Reader, args []string) error {
  72         sf.didBatch = append(sf.didBatch, args)
  73         return sf.errBatch
  74 }
  75
  76 func (sf *slurmFake) QueueCommand(args []string) *exec.Cmd {
  77         return exec.Command("echo", sf.queue)
  78 }
  79
  80 func (sf *slurmFake) Release(name string) error {
  81         sf.didRelease = append(sf.didRelease, name)
  82         return nil
  83 }
  84
  85 func (sf *slurmFake) Renice(name string, nice int64) error {
  86         sf.didRenice = append(sf.didRenice, []string{name, fmt.Sprintf("%d", nice)})
  87         if sf.rejectNice10K && nice > 10000 {
  88                 return errors.New("scontrol: error: Invalid nice value, must be between -10000 and 10000")
  89         }
  90         return nil
  91 }
  92
  93 func (sf *slurmFake) Cancel(name string) error {
  94         sf.didCancel = append(sf.didCancel, name)
  95         if len(sf.didCancel) == 1 {
  96                 // simulate error on first attempt
  97                 return errors.New("something terrible happened")
  98         }
  99         if sf.onCancel != nil {
 100                 sf.onCancel()
 101         }
 102         return nil
 103 }
 104
 105 func (s *IntegrationSuite) integrationTest(c *C,
 106         expectBatch [][]string,
 107         runContainer func(*dispatch.Dispatcher, arvados.Container)) arvados.Container {
 108         arvadostest.ResetEnv()
 109
 110         arv, err := arvadosclient.MakeArvadosClient()
 111         c.Assert(err, IsNil)
 112
 113         // There should be one queued container
 114         params := arvadosclient.Dict{
 115                 "filters": [][]string{{"state", "=", "Queued"}},
 116         }
 117         var containers arvados.ContainerList
 118         err = arv.List("containers", params, &containers)
 119         c.Check(err, IsNil)
 120         c.Assert(len(containers.Items), Equals, 1)
 121
 122         s.disp.cluster.Containers.CrunchRunCommand = "echo"
 123
 124         ctx, cancel := context.WithCancel(context.Background())
 125         doneRun := make(chan struct{})
 126
 127         s.disp.Dispatcher = &dispatch.Dispatcher{
 128                 Arv:        arv,
 129                 PollPeriod: time.Second,
 130                 RunContainer: func(disp *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados.Container) {
 131                         go func() {
 132                                 runContainer(disp, ctr)
 133                                 s.slurm.queue = ""
 134                                 doneRun <- struct{}{}
 135                         }()
 136                         s.disp.runContainer(disp, ctr, status)
 137                         cancel()
 138                 },
 139         }
 140
 141         s.disp.slurm = &s.slurm
 142         s.disp.sqCheck = &SqueueChecker{
 143                 Logger: logrus.StandardLogger(),
 144                 Period: 500 * time.Millisecond,
 145                 Slurm:  s.disp.slurm,
 146         }
 147
 148         err = s.disp.Dispatcher.Run(ctx)
 149         <-doneRun
 150         c.Assert(err, Equals, context.Canceled)
 151
 152         s.disp.sqCheck.Stop()
 153
 154         c.Check(s.slurm.didBatch, DeepEquals, expectBatch)
 155
 156         // There should be no queued containers now
 157         err = arv.List("containers", params, &containers)
 158         c.Check(err, IsNil)
 159         c.Check(len(containers.Items), Equals, 0)
 160
 161         // Previously "Queued" container should now be in "Complete" state
 162         var container arvados.Container
 163         err = arv.Get("containers", "zzzzz-dz642-queuedcontainer", nil, &container)
 164         c.Check(err, IsNil)
 165         return container
 166 }
 167
 168 func (s *IntegrationSuite) TestNormal(c *C) {
 169         s.slurm = slurmFake{queue: "zzzzz-dz642-queuedcontainer 10000 100 PENDING Resources\n"}
 170         container := s.integrationTest(c,
 171                 nil,
 172                 func(dispatcher *dispatch.Dispatcher, container arvados.Container) {
 173                         dispatcher.UpdateState(container.UUID, dispatch.Running)
 174                         time.Sleep(3 * time.Second)
 175                         dispatcher.UpdateState(container.UUID, dispatch.Complete)
 176                 })
 177         c.Check(container.State, Equals, arvados.ContainerStateComplete)
 178 }
 179
 180 func (s *IntegrationSuite) TestCancel(c *C) {
 181         s.slurm = slurmFake{queue: "zzzzz-dz642-queuedcontainer 10000 100 PENDING Resources\n"}
 182         readyToCancel := make(chan bool)
 183         s.slurm.onCancel = func() { <-readyToCancel }
 184         container := s.integrationTest(c,
 185                 nil,
 186                 func(dispatcher *dispatch.Dispatcher, container arvados.Container) {
 187                         dispatcher.UpdateState(container.UUID, dispatch.Running)
 188                         time.Sleep(time.Second)
 189                         dispatcher.Arv.Update("containers", container.UUID,
 190                                 arvadosclient.Dict{
 191                                         "container": arvadosclient.Dict{"priority": 0}},
 192                                 nil)
 193                         readyToCancel <- true
 194                         close(readyToCancel)
 195                 })
 196         c.Check(container.State, Equals, arvados.ContainerStateCancelled)
 197         c.Check(len(s.slurm.didCancel) > 1, Equals, true)
 198         c.Check(s.slurm.didCancel[:2], DeepEquals, []string{"zzzzz-dz642-queuedcontainer", "zzzzz-dz642-queuedcontainer"})
 199 }
 200
 201 func (s *IntegrationSuite) TestMissingFromSqueue(c *C) {
 202         container := s.integrationTest(c,
 203                 [][]string{{
 204                         fmt.Sprintf("--job-name=%s", "zzzzz-dz642-queuedcontainer"),
 205                         fmt.Sprintf("--nice=%d", 10000),
 206                         "--no-requeue",
 207                         fmt.Sprintf("--mem=%d", 11445),
 208                         fmt.Sprintf("--cpus-per-task=%d", 4),
 209                         fmt.Sprintf("--tmp=%d", 45777),
 210                 }},
 211                 func(dispatcher *dispatch.Dispatcher, container arvados.Container) {
 212                         dispatcher.UpdateState(container.UUID, dispatch.Running)
 213                         time.Sleep(3 * time.Second)
 214                         dispatcher.UpdateState(container.UUID, dispatch.Complete)
 215                 })
 216         c.Check(container.State, Equals, arvados.ContainerStateCancelled)
 217 }
 218
 219 func (s *IntegrationSuite) TestSbatchFail(c *C) {
 220         s.slurm = slurmFake{errBatch: errors.New("something terrible happened")}
 221         container := s.integrationTest(c,
 222                 [][]string{{"--job-name=zzzzz-dz642-queuedcontainer", "--nice=10000", "--no-requeue", "--mem=11445", "--cpus-per-task=4", "--tmp=45777"}},
 223                 func(dispatcher *dispatch.Dispatcher, container arvados.Container) {
 224                         dispatcher.UpdateState(container.UUID, dispatch.Running)
 225                         dispatcher.UpdateState(container.UUID, dispatch.Complete)
 226                 })
 227         c.Check(container.State, Equals, arvados.ContainerStateComplete)
 228
 229         arv, err := arvadosclient.MakeArvadosClient()
 230         c.Assert(err, IsNil)
 231
 232         var ll arvados.LogList
 233         err = arv.List("logs", arvadosclient.Dict{"filters": [][]string{
 234                 {"object_uuid", "=", container.UUID},
 235                 {"event_type", "=", "dispatch"},
 236         }}, &ll)
 237         c.Assert(err, IsNil)
 238         c.Assert(len(ll.Items), Equals, 1)
 239 }
 240
 241 type StubbedSuite struct {
 242         disp Dispatcher
 243 }
 244
 245 func (s *StubbedSuite) SetUpTest(c *C) {
 246         s.disp = Dispatcher{}
 247         s.disp.cluster = &arvados.Cluster{}
 248         s.disp.setup()
 249 }
 250
 251 func (s *StubbedSuite) TestAPIErrorGettingContainers(c *C) {
 252         apiStubResponses := make(map[string]arvadostest.StubResponse)
 253         apiStubResponses["/arvados/v1/api_client_authorizations/current"] = arvadostest.StubResponse{200, `{"uuid":"` + arvadostest.Dispatch1AuthUUID + `"}`}
 254         apiStubResponses["/arvados/v1/containers"] = arvadostest.StubResponse{500, string(`{}`)}
 255
 256         s.testWithServerStub(c, apiStubResponses, "echo", "error getting count of containers")
 257 }
 258
 259 func (s *StubbedSuite) testWithServerStub(c *C, apiStubResponses map[string]arvadostest.StubResponse, crunchCmd string, expected string) {
 260         apiStub := arvadostest.ServerStub{apiStubResponses}
 261
 262         api := httptest.NewServer(&apiStub)
 263         defer api.Close()
 264
 265         arv := &arvadosclient.ArvadosClient{
 266                 Scheme:    "http",
 267                 ApiServer: api.URL[7:],
 268                 ApiToken:  "abc123",
 269                 Client:    &http.Client{Transport: &http.Transport{}},
 270                 Retries:   0,
 271         }
 272
 273         buf := bytes.NewBuffer(nil)
 274         logrus.SetOutput(io.MultiWriter(buf, os.Stderr))
 275         defer logrus.SetOutput(os.Stderr)
 276
 277         s.disp.cluster.Containers.CrunchRunCommand = "crunchCmd"
 278
 279         ctx, cancel := context.WithCancel(context.Background())
 280         dispatcher := dispatch.Dispatcher{
 281                 Arv:        arv,
 282                 PollPeriod: time.Second,
 283                 RunContainer: func(disp *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados.Container) {
 284                         go func() {
 285                                 time.Sleep(time.Second)
 286                                 disp.UpdateState(ctr.UUID, dispatch.Running)
 287                                 disp.UpdateState(ctr.UUID, dispatch.Complete)
 288                         }()
 289                         s.disp.runContainer(disp, ctr, status)
 290                         cancel()
 291                 },
 292         }
 293
 294         go func() {
 295                 for i := 0; i < 80 && !strings.Contains(buf.String(), expected); i++ {
 296                         time.Sleep(100 * time.Millisecond)
 297                 }
 298                 cancel()
 299         }()
 300
 301         err := dispatcher.Run(ctx)
 302         c.Assert(err, Equals, context.Canceled)
 303
 304         c.Check(buf.String(), Matches, `(?ms).*`+expected+`.*`)
 305 }
 306
 307 func (s *StubbedSuite) TestSbatchArgs(c *C) {
 308         container := arvados.Container{
 309                 UUID:               "123",
 310                 RuntimeConstraints: arvados.RuntimeConstraints{RAM: 250000000, VCPUs: 2},
 311                 Priority:           1,
 312         }
 313
 314         for _, defaults := range [][]string{
 315                 nil,
 316                 {},
 317                 {"--arg1=v1", "--arg2"},
 318         } {
 319                 c.Logf("%#v", defaults)
 320                 s.disp.cluster.Containers.SLURM.SbatchArgumentsList = defaults
 321
 322                 args, err := s.disp.sbatchArgs(container)
 323                 c.Check(args, DeepEquals, append(defaults, "--job-name=123", "--nice=10000", "--no-requeue", "--mem=239", "--cpus-per-task=2", "--tmp=0"))
 324                 c.Check(err, IsNil)
 325         }
 326 }
 327
 328 func (s *StubbedSuite) TestSbatchInstanceTypeConstraint(c *C) {
 329         container := arvados.Container{
 330                 UUID:               "123",
 331                 RuntimeConstraints: arvados.RuntimeConstraints{RAM: 250000000, VCPUs: 2},
 332                 Priority:           1,
 333         }
 334
 335         for _, trial := range []struct {
 336                 types      map[string]arvados.InstanceType
 337                 sbatchArgs []string
 338                 err        error
 339         }{
 340                 // Choose node type => use --constraint arg
 341                 {
 342                         types: map[string]arvados.InstanceType{
 343                                 "a1.tiny":   {Name: "a1.tiny", Price: 0.02, RAM: 128000000, VCPUs: 1},
 344                                 "a1.small":  {Name: "a1.small", Price: 0.04, RAM: 256000000, VCPUs: 2},
 345                                 "a1.medium": {Name: "a1.medium", Price: 0.08, RAM: 512000000, VCPUs: 4},
 346                                 "a1.large":  {Name: "a1.large", Price: 0.16, RAM: 1024000000, VCPUs: 8},
 347                         },
 348                         sbatchArgs: []string{"--constraint=instancetype=a1.medium"},
 349                 },
 350                 // No node types configured => no slurm constraint
 351                 {
 352                         types:      nil,
 353                         sbatchArgs: []string{"--mem=239", "--cpus-per-task=2", "--tmp=0"},
 354                 },
 355                 // No node type is big enough => error
 356                 {
 357                         types: map[string]arvados.InstanceType{
 358                                 "a1.tiny": {Name: "a1.tiny", Price: 0.02, RAM: 128000000, VCPUs: 1},
 359                         },
 360                         err: dispatchcloud.ConstraintsNotSatisfiableError{},
 361                 },
 362         } {
 363                 c.Logf("%#v", trial)
 364                 s.disp.cluster = &arvados.Cluster{InstanceTypes: trial.types}
 365
 366                 args, err := s.disp.sbatchArgs(container)
 367                 c.Check(err == nil, Equals, trial.err == nil)
 368                 if trial.err == nil {
 369                         c.Check(args, DeepEquals, append([]string{"--job-name=123", "--nice=10000", "--no-requeue"}, trial.sbatchArgs...))
 370                 } else {
 371                         c.Check(len(err.(dispatchcloud.ConstraintsNotSatisfiableError).AvailableTypes), Equals, len(trial.types))
 372                 }
 373         }
 374 }
 375
 376 func (s *StubbedSuite) TestSbatchPartition(c *C) {
 377         container := arvados.Container{
 378                 UUID:                 "123",
 379                 RuntimeConstraints:   arvados.RuntimeConstraints{RAM: 250000000, VCPUs: 1},
 380                 SchedulingParameters: arvados.SchedulingParameters{Partitions: []string{"blurb", "b2"}},
 381                 Priority:             1,
 382         }
 383
 384         args, err := s.disp.sbatchArgs(container)
 385         c.Check(args, DeepEquals, []string{
 386                 "--job-name=123", "--nice=10000", "--no-requeue",
 387                 "--mem=239", "--cpus-per-task=1", "--tmp=0",
 388                 "--partition=blurb,b2",
 389         })
 390         c.Check(err, IsNil)
 391 }
 392
 393 func (s *StubbedSuite) TestLoadLegacyConfig(c *C) {
 394         content := []byte(`
 395 Client:
 396   APIHost: example.com
 397   AuthToken: abcdefg
 398   KeepServiceURIs:
 399     - https://example.com/keep1
 400     - https://example.com/keep2
 401 SbatchArguments: ["--foo", "bar"]
 402 PollPeriod: 12s
 403 PrioritySpread: 42
 404 CrunchRunCommand: ["x-crunch-run", "--cgroup-parent-subsystem=memory"]
 405 ReserveExtraRAM: 12345
 406 MinRetryPeriod: 13s
 407 BatchSize: 99
 408 `)
 409         tmpfile, err := ioutil.TempFile("", "example")
 410         if err != nil {
 411                 c.Error(err)
 412         }
 413
 414         defer os.Remove(tmpfile.Name()) // clean up
 415
 416         if _, err := tmpfile.Write(content); err != nil {
 417                 c.Error(err)
 418         }
 419         if err := tmpfile.Close(); err != nil {
 420                 c.Error(err)
 421
 422         }
 423         err = s.disp.configure("crunch-dispatch-slurm", []string{"-config", tmpfile.Name()})
 424         c.Check(err, IsNil)
 425
 426         c.Check(s.disp.cluster.Services.Controller.ExternalURL, Equals, arvados.URL{Scheme: "https", Host: "example.com"})
 427         c.Check(s.disp.cluster.SystemRootToken, Equals, "abcdefg")
 428         c.Check(s.disp.cluster.Containers.SLURM.SbatchArgumentsList, DeepEquals, []string{"--foo", "bar"})
 429         c.Check(s.disp.cluster.Containers.CloudVMs.PollInterval, Equals, arvados.Duration(12*time.Second))
 430         c.Check(s.disp.cluster.Containers.SLURM.PrioritySpread, Equals, int64(42))
 431         c.Check(s.disp.cluster.Containers.CrunchRunCommand, Equals, "x-crunch-run")
 432         c.Check(s.disp.cluster.Containers.CrunchRunArgumentsList, DeepEquals, []string{"--cgroup-parent-subsystem=memory"})
 433         c.Check(s.disp.cluster.Containers.ReserveExtraRAM, Equals, arvados.ByteSize(12345))
 434         c.Check(s.disp.cluster.Containers.MinRetryPeriod, Equals, arvados.Duration(13*time.Second))
 435         c.Check(s.disp.cluster.API.MaxItemsPerResponse, Equals, 99)
 436         c.Check(s.disp.cluster.Containers.SLURM.KeepServices, DeepEquals, map[string]arvados.Service{
 437                 "00000-bi6l4-000000000000000": arvados.Service{
 438                         InternalURLs: map[arvados.URL]arvados.ServiceInstance{
 439                                 arvados.URL{Scheme: "https", Path: "/keep1", Host: "example.com"}: struct{}{},
 440                                 arvados.URL{Scheme: "https", Path: "/keep2", Host: "example.com"}: struct{}{},
 441                         },
 442                 },
 443         })
 444         ks := os.Getenv("ARVADOS_KEEP_SERVICES")
 445         if ks != "https://example.com/keep1 https://example.com/keep2" && ks != "https://example.com/keep2 https://example.com/keep1" {
 446                 c.Assert(ks, Equals, "https://example.com/keep1 https://example.com/keep2")
 447         }
 448 }