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         "log"
  15         "net/http"
  16         "net/http/httptest"
  17         "os"
  18         "os/exec"
  19         "strings"
  20         "testing"
  21         "time"
  22
  23         "git.curoverse.com/arvados.git/lib/dispatchcloud"
  24         "git.curoverse.com/arvados.git/sdk/go/arvados"
  25         "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
  26         "git.curoverse.com/arvados.git/sdk/go/arvadostest"
  27         "git.curoverse.com/arvados.git/sdk/go/dispatch"
  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.setup()
  49         s.slurm = slurmFake{}
  50 }
  51
  52 func (s *IntegrationSuite) TearDownTest(c *C) {
  53         arvadostest.ResetEnv()
  54         arvadostest.StopAPI()
  55 }
  56
  57 type slurmFake struct {
  58         didBatch   [][]string
  59         didCancel  []string
  60         didRelease []string
  61         didRenice  [][]string
  62         queue      string
  63         // If non-nil, run this func during the 2nd+ call to Cancel()
  64         onCancel func()
  65         // Error returned by Batch()
  66         errBatch error
  67 }
  68
  69 func (sf *slurmFake) Batch(script io.Reader, args []string) error {
  70         sf.didBatch = append(sf.didBatch, args)
  71         return sf.errBatch
  72 }
  73
  74 func (sf *slurmFake) QueueCommand(args []string) *exec.Cmd {
  75         return exec.Command("echo", sf.queue)
  76 }
  77
  78 func (sf *slurmFake) Release(name string) error {
  79         sf.didRelease = append(sf.didRelease, name)
  80         return nil
  81 }
  82
  83 func (sf *slurmFake) Renice(name string, nice int64) error {
  84         sf.didRenice = append(sf.didRenice, []string{name, fmt.Sprintf("%d", nice)})
  85         return nil
  86 }
  87
  88 func (sf *slurmFake) Cancel(name string) error {
  89         sf.didCancel = append(sf.didCancel, name)
  90         if len(sf.didCancel) == 1 {
  91                 // simulate error on first attempt
  92                 return errors.New("something terrible happened")
  93         }
  94         if sf.onCancel != nil {
  95                 sf.onCancel()
  96         }
  97         return nil
  98 }
  99
 100 func (s *IntegrationSuite) integrationTest(c *C,
 101         expectBatch [][]string,
 102         runContainer func(*dispatch.Dispatcher, arvados.Container)) arvados.Container {
 103         arvadostest.ResetEnv()
 104
 105         arv, err := arvadosclient.MakeArvadosClient()
 106         c.Assert(err, IsNil)
 107
 108         // There should be one queued container
 109         params := arvadosclient.Dict{
 110                 "filters": [][]string{{"state", "=", "Queued"}},
 111         }
 112         var containers arvados.ContainerList
 113         err = arv.List("containers", params, &containers)
 114         c.Check(err, IsNil)
 115         c.Check(len(containers.Items), Equals, 1)
 116
 117         s.disp.CrunchRunCommand = []string{"echo"}
 118
 119         ctx, cancel := context.WithCancel(context.Background())
 120         doneRun := make(chan struct{})
 121
 122         s.disp.Dispatcher = &dispatch.Dispatcher{
 123                 Arv:        arv,
 124                 PollPeriod: time.Second,
 125                 RunContainer: func(disp *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados.Container) {
 126                         go func() {
 127                                 runContainer(disp, ctr)
 128                                 s.slurm.queue = ""
 129                                 doneRun <- struct{}{}
 130                         }()
 131                         s.disp.runContainer(disp, ctr, status)
 132                         cancel()
 133                 },
 134         }
 135
 136         s.disp.slurm = &s.slurm
 137         s.disp.sqCheck = &SqueueChecker{Period: 500 * time.Millisecond, Slurm: s.disp.slurm}
 138
 139         err = s.disp.Dispatcher.Run(ctx)
 140         <-doneRun
 141         c.Assert(err, Equals, context.Canceled)
 142
 143         s.disp.sqCheck.Stop()
 144
 145         c.Check(s.slurm.didBatch, DeepEquals, expectBatch)
 146
 147         // There should be no queued containers now
 148         err = arv.List("containers", params, &containers)
 149         c.Check(err, IsNil)
 150         c.Check(len(containers.Items), Equals, 0)
 151
 152         // Previously "Queued" container should now be in "Complete" state
 153         var container arvados.Container
 154         err = arv.Get("containers", "zzzzz-dz642-queuedcontainer", nil, &container)
 155         c.Check(err, IsNil)
 156         return container
 157 }
 158
 159 func (s *IntegrationSuite) TestNormal(c *C) {
 160         s.slurm = slurmFake{queue: "zzzzz-dz642-queuedcontainer 10000 100 PENDING Resources\n"}
 161         container := s.integrationTest(c,
 162                 nil,
 163                 func(dispatcher *dispatch.Dispatcher, container arvados.Container) {
 164                         dispatcher.UpdateState(container.UUID, dispatch.Running)
 165                         time.Sleep(3 * time.Second)
 166                         dispatcher.UpdateState(container.UUID, dispatch.Complete)
 167                 })
 168         c.Check(container.State, Equals, arvados.ContainerStateComplete)
 169 }
 170
 171 func (s *IntegrationSuite) TestCancel(c *C) {
 172         s.slurm = slurmFake{queue: "zzzzz-dz642-queuedcontainer 10000 100 PENDING Resources\n"}
 173         readyToCancel := make(chan bool)
 174         s.slurm.onCancel = func() { <-readyToCancel }
 175         container := s.integrationTest(c,
 176                 nil,
 177                 func(dispatcher *dispatch.Dispatcher, container arvados.Container) {
 178                         dispatcher.UpdateState(container.UUID, dispatch.Running)
 179                         time.Sleep(time.Second)
 180                         dispatcher.Arv.Update("containers", container.UUID,
 181                                 arvadosclient.Dict{
 182                                         "container": arvadosclient.Dict{"priority": 0}},
 183                                 nil)
 184                         readyToCancel <- true
 185                         close(readyToCancel)
 186                 })
 187         c.Check(container.State, Equals, arvados.ContainerStateCancelled)
 188         c.Check(len(s.slurm.didCancel) > 1, Equals, true)
 189         c.Check(s.slurm.didCancel[:2], DeepEquals, []string{"zzzzz-dz642-queuedcontainer", "zzzzz-dz642-queuedcontainer"})
 190 }
 191
 192 func (s *IntegrationSuite) TestMissingFromSqueue(c *C) {
 193         container := s.integrationTest(c,
 194                 [][]string{{
 195                         fmt.Sprintf("--job-name=%s", "zzzzz-dz642-queuedcontainer"),
 196                         fmt.Sprintf("--mem=%d", 11445),
 197                         fmt.Sprintf("--cpus-per-task=%d", 4),
 198                         fmt.Sprintf("--tmp=%d", 45777),
 199                         fmt.Sprintf("--nice=%d", 10000)}},
 200                 func(dispatcher *dispatch.Dispatcher, container arvados.Container) {
 201                         dispatcher.UpdateState(container.UUID, dispatch.Running)
 202                         time.Sleep(3 * time.Second)
 203                         dispatcher.UpdateState(container.UUID, dispatch.Complete)
 204                 })
 205         c.Check(container.State, Equals, arvados.ContainerStateCancelled)
 206 }
 207
 208 func (s *IntegrationSuite) TestSbatchFail(c *C) {
 209         s.slurm = slurmFake{errBatch: errors.New("something terrible happened")}
 210         container := s.integrationTest(c,
 211                 [][]string{{"--job-name=zzzzz-dz642-queuedcontainer", "--mem=11445", "--cpus-per-task=4", "--tmp=45777", "--nice=10000"}},
 212                 func(dispatcher *dispatch.Dispatcher, container arvados.Container) {
 213                         dispatcher.UpdateState(container.UUID, dispatch.Running)
 214                         dispatcher.UpdateState(container.UUID, dispatch.Complete)
 215                 })
 216         c.Check(container.State, Equals, arvados.ContainerStateComplete)
 217
 218         arv, err := arvadosclient.MakeArvadosClient()
 219         c.Assert(err, IsNil)
 220
 221         var ll arvados.LogList
 222         err = arv.List("logs", arvadosclient.Dict{"filters": [][]string{
 223                 {"object_uuid", "=", container.UUID},
 224                 {"event_type", "=", "dispatch"},
 225         }}, &ll)
 226         c.Assert(err, IsNil)
 227         c.Assert(len(ll.Items), Equals, 1)
 228 }
 229
 230 type StubbedSuite struct {
 231         disp Dispatcher
 232 }
 233
 234 func (s *StubbedSuite) SetUpTest(c *C) {
 235         s.disp = Dispatcher{}
 236         s.disp.setup()
 237 }
 238
 239 func (s *StubbedSuite) TestAPIErrorGettingContainers(c *C) {
 240         apiStubResponses := make(map[string]arvadostest.StubResponse)
 241         apiStubResponses["/arvados/v1/api_client_authorizations/current"] = arvadostest.StubResponse{200, `{"uuid":"` + arvadostest.Dispatch1AuthUUID + `"}`}
 242         apiStubResponses["/arvados/v1/containers"] = arvadostest.StubResponse{500, string(`{}`)}
 243
 244         s.testWithServerStub(c, apiStubResponses, "echo", "Error getting list of containers")
 245 }
 246
 247 func (s *StubbedSuite) testWithServerStub(c *C, apiStubResponses map[string]arvadostest.StubResponse, crunchCmd string, expected string) {
 248         apiStub := arvadostest.ServerStub{apiStubResponses}
 249
 250         api := httptest.NewServer(&apiStub)
 251         defer api.Close()
 252
 253         arv := &arvadosclient.ArvadosClient{
 254                 Scheme:    "http",
 255                 ApiServer: api.URL[7:],
 256                 ApiToken:  "abc123",
 257                 Client:    &http.Client{Transport: &http.Transport{}},
 258                 Retries:   0,
 259         }
 260
 261         buf := bytes.NewBuffer(nil)
 262         log.SetOutput(io.MultiWriter(buf, os.Stderr))
 263         defer log.SetOutput(os.Stderr)
 264
 265         s.disp.CrunchRunCommand = []string{crunchCmd}
 266
 267         ctx, cancel := context.WithCancel(context.Background())
 268         dispatcher := dispatch.Dispatcher{
 269                 Arv:        arv,
 270                 PollPeriod: time.Second,
 271                 RunContainer: func(disp *dispatch.Dispatcher, ctr arvados.Container, status <-chan arvados.Container) {
 272                         go func() {
 273                                 time.Sleep(time.Second)
 274                                 disp.UpdateState(ctr.UUID, dispatch.Running)
 275                                 disp.UpdateState(ctr.UUID, dispatch.Complete)
 276                         }()
 277                         s.disp.runContainer(disp, ctr, status)
 278                         cancel()
 279                 },
 280         }
 281
 282         go func() {
 283                 for i := 0; i < 80 && !strings.Contains(buf.String(), expected); i++ {
 284                         time.Sleep(100 * time.Millisecond)
 285                 }
 286                 cancel()
 287         }()
 288
 289         err := dispatcher.Run(ctx)
 290         c.Assert(err, Equals, context.Canceled)
 291
 292         c.Check(buf.String(), Matches, `(?ms).*`+expected+`.*`)
 293 }
 294
 295 func (s *StubbedSuite) TestNoSuchConfigFile(c *C) {
 296         err := s.disp.readConfig("/nosuchdir89j7879/8hjwr7ojgyy7")
 297         c.Assert(err, NotNil)
 298 }
 299
 300 func (s *StubbedSuite) TestBadSbatchArgsConfig(c *C) {
 301         tmpfile, err := ioutil.TempFile(os.TempDir(), "config")
 302         c.Check(err, IsNil)
 303         defer os.Remove(tmpfile.Name())
 304
 305         _, err = tmpfile.Write([]byte(`{"SbatchArguments": "oops this is not a string array"}`))
 306         c.Check(err, IsNil)
 307
 308         err = s.disp.readConfig(tmpfile.Name())
 309         c.Assert(err, NotNil)
 310 }
 311
 312 func (s *StubbedSuite) TestNoSuchArgInConfigIgnored(c *C) {
 313         tmpfile, err := ioutil.TempFile(os.TempDir(), "config")
 314         c.Check(err, IsNil)
 315         defer os.Remove(tmpfile.Name())
 316
 317         _, err = tmpfile.Write([]byte(`{"NoSuchArg": "Nobody loves me, not one tiny hunk."}`))
 318         c.Check(err, IsNil)
 319
 320         err = s.disp.readConfig(tmpfile.Name())
 321         c.Assert(err, IsNil)
 322         c.Check(0, Equals, len(s.disp.SbatchArguments))
 323 }
 324
 325 func (s *StubbedSuite) TestReadConfig(c *C) {
 326         tmpfile, err := ioutil.TempFile(os.TempDir(), "config")
 327         c.Check(err, IsNil)
 328         defer os.Remove(tmpfile.Name())
 329
 330         args := []string{"--arg1=v1", "--arg2", "--arg3=v3"}
 331         argsS := `{"SbatchArguments": ["--arg1=v1",  "--arg2", "--arg3=v3"]}`
 332         _, err = tmpfile.Write([]byte(argsS))
 333         c.Check(err, IsNil)
 334
 335         err = s.disp.readConfig(tmpfile.Name())
 336         c.Assert(err, IsNil)
 337         c.Check(args, DeepEquals, s.disp.SbatchArguments)
 338 }
 339
 340 func (s *StubbedSuite) TestSbatchArgs(c *C) {
 341         container := arvados.Container{
 342                 UUID:               "123",
 343                 RuntimeConstraints: arvados.RuntimeConstraints{RAM: 250000000, VCPUs: 2},
 344                 Priority:           1,
 345         }
 346
 347         for _, defaults := range [][]string{
 348                 nil,
 349                 {},
 350                 {"--arg1=v1", "--arg2"},
 351         } {
 352                 c.Logf("%#v", defaults)
 353                 s.disp.SbatchArguments = defaults
 354
 355                 args, err := s.disp.sbatchArgs(container)
 356                 c.Check(args, DeepEquals, append(defaults, "--job-name=123", "--mem=239", "--cpus-per-task=2", "--tmp=0", "--nice=10000"))
 357                 c.Check(err, IsNil)
 358         }
 359 }
 360
 361 func (s *StubbedSuite) TestSbatchInstanceTypeConstraint(c *C) {
 362         container := arvados.Container{
 363                 UUID:               "123",
 364                 RuntimeConstraints: arvados.RuntimeConstraints{RAM: 250000000, VCPUs: 2},
 365                 Priority:           1,
 366         }
 367
 368         for _, trial := range []struct {
 369                 types      []arvados.InstanceType
 370                 sbatchArgs []string
 371                 err        error
 372         }{
 373                 // Choose node type => use --constraint arg
 374                 {
 375                         types: []arvados.InstanceType{
 376                                 {Name: "a1.tiny", Price: 0.02, RAM: 128000000, VCPUs: 1},
 377                                 {Name: "a1.small", Price: 0.04, RAM: 256000000, VCPUs: 2},
 378                                 {Name: "a1.medium", Price: 0.08, RAM: 512000000, VCPUs: 4},
 379                                 {Name: "a1.large", Price: 0.16, RAM: 1024000000, VCPUs: 8},
 380                         },
 381                         sbatchArgs: []string{"--constraint=instancetype=a1.medium"},
 382                 },
 383                 // No node types configured => no slurm constraint
 384                 {
 385                         types:      nil,
 386                         sbatchArgs: nil,
 387                 },
 388                 // No node type is big enough => error
 389                 {
 390                         types: []arvados.InstanceType{
 391                                 {Name: "a1.tiny", Price: 0.02, RAM: 128000000, VCPUs: 1},
 392                         },
 393                         err: dispatchcloud.ErrConstraintsNotSatisfiable,
 394                 },
 395         } {
 396                 c.Logf("%#v", trial)
 397                 s.disp.cluster = &arvados.Cluster{InstanceTypes: trial.types}
 398
 399                 args, err := s.disp.sbatchArgs(container)
 400                 c.Check(err, Equals, trial.err)
 401                 if trial.err == nil {
 402                         c.Check(args, DeepEquals, append([]string{"--job-name=123", "--mem=239", "--cpus-per-task=2", "--tmp=0", "--nice=10000"}, trial.sbatchArgs...))
 403                 }
 404         }
 405 }
 406
 407 func (s *StubbedSuite) TestSbatchPartition(c *C) {
 408         container := arvados.Container{
 409                 UUID:                 "123",
 410                 RuntimeConstraints:   arvados.RuntimeConstraints{RAM: 250000000, VCPUs: 1},
 411                 SchedulingParameters: arvados.SchedulingParameters{Partitions: []string{"blurb", "b2"}},
 412                 Priority:             1,
 413         }
 414
 415         args, err := s.disp.sbatchArgs(container)
 416         c.Check(args, DeepEquals, []string{
 417                 "--job-name=123", "--mem=239", "--cpus-per-task=1", "--tmp=0", "--nice=10000",
 418                 "--partition=blurb,b2",
 419         })
 420         c.Check(err, IsNil)
 421 }