lib/lsf/dispatch_test.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package lsf
   6
   7 import (
   8         "context"
   9         "encoding/json"
  10         "fmt"
  11         "math/rand"
  12         "os/exec"
  13         "strconv"
  14         "sync"
  15         "testing"
  16         "time"
  17
  18         "git.arvados.org/arvados.git/lib/config"
  19         "git.arvados.org/arvados.git/sdk/go/arvados"
  20         "git.arvados.org/arvados.git/sdk/go/arvadostest"
  21         "git.arvados.org/arvados.git/sdk/go/ctxlog"
  22         "github.com/prometheus/client_golang/prometheus"
  23         "gopkg.in/check.v1"
  24 )
  25
  26 func Test(t *testing.T) {
  27         check.TestingT(t)
  28 }
  29
  30 var _ = check.Suite(&suite{})
  31
  32 type suite struct {
  33         disp          *dispatcher
  34         crTooBig      arvados.ContainerRequest
  35         crCUDARequest arvados.ContainerRequest
  36 }
  37
  38 func (s *suite) TearDownTest(c *check.C) {
  39         arvados.NewClientFromEnv().RequestAndDecode(nil, "POST", "database/reset", nil, nil)
  40 }
  41
  42 func (s *suite) SetUpTest(c *check.C) {
  43         cfg, err := config.NewLoader(nil, ctxlog.TestLogger(c)).Load()
  44         c.Assert(err, check.IsNil)
  45         cluster, err := cfg.GetCluster("")
  46         c.Assert(err, check.IsNil)
  47         cluster.Containers.CloudVMs.PollInterval = arvados.Duration(time.Second)
  48         s.disp = newHandler(context.Background(), cluster, arvadostest.Dispatch1Token, prometheus.NewRegistry()).(*dispatcher)
  49         s.disp.lsfcli.stubCommand = func(string, ...string) *exec.Cmd {
  50                 return exec.Command("bash", "-c", "echo >&2 unimplemented stub; false")
  51         }
  52         err = arvados.NewClientFromEnv().RequestAndDecode(&s.crTooBig, "POST", "arvados/v1/container_requests", nil, map[string]interface{}{
  53                 "container_request": map[string]interface{}{
  54                         "runtime_constraints": arvados.RuntimeConstraints{
  55                                 RAM:   1000000000000,
  56                                 VCPUs: 1,
  57                         },
  58                         "container_image":     arvadostest.DockerImage112PDH,
  59                         "command":             []string{"sleep", "1"},
  60                         "mounts":              map[string]arvados.Mount{"/mnt/out": {Kind: "tmp", Capacity: 1000}},
  61                         "output_path":         "/mnt/out",
  62                         "state":               arvados.ContainerRequestStateCommitted,
  63                         "priority":            1,
  64                         "container_count_max": 1,
  65                 },
  66         })
  67         c.Assert(err, check.IsNil)
  68
  69         err = arvados.NewClientFromEnv().RequestAndDecode(&s.crCUDARequest, "POST", "arvados/v1/container_requests", nil, map[string]interface{}{
  70                 "container_request": map[string]interface{}{
  71                         "runtime_constraints": arvados.RuntimeConstraints{
  72                                 RAM:   16000000,
  73                                 VCPUs: 1,
  74                                 CUDA: arvados.CUDARuntimeConstraints{
  75                                         DeviceCount:        1,
  76                                         DriverVersion:      "11.0",
  77                                         HardwareCapability: "8.0",
  78                                 },
  79                         },
  80                         "container_image":     arvadostest.DockerImage112PDH,
  81                         "command":             []string{"sleep", "1"},
  82                         "mounts":              map[string]arvados.Mount{"/mnt/out": {Kind: "tmp", Capacity: 1000}},
  83                         "output_path":         "/mnt/out",
  84                         "state":               arvados.ContainerRequestStateCommitted,
  85                         "priority":            1,
  86                         "container_count_max": 1,
  87                 },
  88         })
  89         c.Assert(err, check.IsNil)
  90
  91 }
  92
  93 type lsfstub struct {
  94         sudoUser  string
  95         errorRate float64
  96 }
  97
  98 func (stub lsfstub) stubCommand(s *suite, c *check.C) func(prog string, args ...string) *exec.Cmd {
  99         mtx := sync.Mutex{}
 100         nextjobid := 100
 101         fakejobq := map[int]string{}
 102         return func(prog string, args ...string) *exec.Cmd {
 103                 c.Logf("stubCommand: %q %q", prog, args)
 104                 if rand.Float64() < stub.errorRate {
 105                         return exec.Command("bash", "-c", "echo >&2 'stub random failure' && false")
 106                 }
 107                 if stub.sudoUser != "" && len(args) > 3 &&
 108                         prog == "sudo" &&
 109                         args[0] == "-E" &&
 110                         args[1] == "-u" &&
 111                         args[2] == stub.sudoUser {
 112                         prog, args = args[3], args[4:]
 113                 }
 114                 switch prog {
 115                 case "bsub":
 116                         defaultArgs := s.disp.Cluster.Containers.LSF.BsubArgumentsList
 117                         if args[5] == s.crCUDARequest.ContainerUUID {
 118                                 c.Assert(len(args), check.Equals, len(defaultArgs)+len(s.disp.Cluster.Containers.LSF.BsubCUDAArguments))
 119                         } else {
 120                                 c.Assert(len(args), check.Equals, len(defaultArgs))
 121                         }
 122                         // %%J must have been rewritten to %J
 123                         c.Check(args[1], check.Equals, "/tmp/crunch-run.%J.out")
 124                         args = args[4:]
 125                         switch args[1] {
 126                         case arvadostest.LockedContainerUUID:
 127                                 c.Check(args, check.DeepEquals, []string{
 128                                         "-J", arvadostest.LockedContainerUUID,
 129                                         "-n", "4",
 130                                         "-D", "11701MB",
 131                                         "-R", "rusage[mem=11701MB:tmp=0MB] span[hosts=1]",
 132                                         "-R", "select[mem>=11701MB]",
 133                                         "-R", "select[tmp>=0MB]",
 134                                         "-R", "select[ncpus>=4]"})
 135                                 mtx.Lock()
 136                                 fakejobq[nextjobid] = args[1]
 137                                 nextjobid++
 138                                 mtx.Unlock()
 139                         case arvadostest.QueuedContainerUUID:
 140                                 c.Check(args, check.DeepEquals, []string{
 141                                         "-J", arvadostest.QueuedContainerUUID,
 142                                         "-n", "4",
 143                                         "-D", "11701MB",
 144                                         "-R", "rusage[mem=11701MB:tmp=45777MB] span[hosts=1]",
 145                                         "-R", "select[mem>=11701MB]",
 146                                         "-R", "select[tmp>=45777MB]",
 147                                         "-R", "select[ncpus>=4]"})
 148                                 mtx.Lock()
 149                                 fakejobq[nextjobid] = args[1]
 150                                 nextjobid++
 151                                 mtx.Unlock()
 152                         case s.crTooBig.ContainerUUID:
 153                                 c.Check(args, check.DeepEquals, []string{
 154                                         "-J", s.crTooBig.ContainerUUID,
 155                                         "-n", "1",
 156                                         "-D", "954187MB",
 157                                         "-R", "rusage[mem=954187MB:tmp=256MB] span[hosts=1]",
 158                                         "-R", "select[mem>=954187MB]",
 159                                         "-R", "select[tmp>=256MB]",
 160                                         "-R", "select[ncpus>=1]"})
 161                                 mtx.Lock()
 162                                 fakejobq[nextjobid] = args[1]
 163                                 nextjobid++
 164                                 mtx.Unlock()
 165                         case s.crCUDARequest.ContainerUUID:
 166                                 c.Check(args, check.DeepEquals, []string{
 167                                         "-J", s.crCUDARequest.ContainerUUID,
 168                                         "-n", "1",
 169                                         "-D", "528MB",
 170                                         "-R", "rusage[mem=528MB:tmp=256MB] span[hosts=1]",
 171                                         "-R", "select[mem>=528MB]",
 172                                         "-R", "select[tmp>=256MB]",
 173                                         "-R", "select[ncpus>=1]",
 174                                         "-gpu", "num=1"})
 175                                 mtx.Lock()
 176                                 fakejobq[nextjobid] = args[1]
 177                                 nextjobid++
 178                                 mtx.Unlock()
 179                         default:
 180                                 c.Errorf("unexpected uuid passed to bsub: args %q", args)
 181                                 return exec.Command("false")
 182                         }
 183                         return exec.Command("echo", "submitted job")
 184                 case "bjobs":
 185                         c.Check(args, check.DeepEquals, []string{"-u", "all", "-o", "jobid stat job_name pend_reason", "-json"})
 186                         var records []map[string]interface{}
 187                         for jobid, uuid := range fakejobq {
 188                                 stat, reason := "RUN", ""
 189                                 if uuid == s.crTooBig.ContainerUUID {
 190                                         // The real bjobs output includes a trailing ';' here:
 191                                         stat, reason = "PEND", "There are no suitable hosts for the job;"
 192                                 }
 193                                 records = append(records, map[string]interface{}{
 194                                         "JOBID":       fmt.Sprintf("%d", jobid),
 195                                         "STAT":        stat,
 196                                         "JOB_NAME":    uuid,
 197                                         "PEND_REASON": reason,
 198                                 })
 199                         }
 200                         out, err := json.Marshal(map[string]interface{}{
 201                                 "COMMAND": "bjobs",
 202                                 "JOBS":    len(fakejobq),
 203                                 "RECORDS": records,
 204                         })
 205                         if err != nil {
 206                                 panic(err)
 207                         }
 208                         c.Logf("bjobs out: %s", out)
 209                         return exec.Command("printf", string(out))
 210                 case "bkill":
 211                         killid, _ := strconv.Atoi(args[0])
 212                         if uuid, ok := fakejobq[killid]; !ok {
 213                                 return exec.Command("bash", "-c", fmt.Sprintf("printf >&2 'Job <%d>: No matching job found\n'", killid))
 214                         } else if uuid == "" {
 215                                 return exec.Command("bash", "-c", fmt.Sprintf("printf >&2 'Job <%d>: Job has already finished\n'", killid))
 216                         } else {
 217                                 go func() {
 218                                         time.Sleep(time.Millisecond)
 219                                         mtx.Lock()
 220                                         delete(fakejobq, killid)
 221                                         mtx.Unlock()
 222                                 }()
 223                                 return exec.Command("bash", "-c", fmt.Sprintf("printf 'Job <%d> is being terminated\n'", killid))
 224                         }
 225                 default:
 226                         return exec.Command("bash", "-c", fmt.Sprintf("echo >&2 'stub: command not found: %+q'", prog))
 227                 }
 228         }
 229 }
 230
 231 func (s *suite) TestSubmit(c *check.C) {
 232         s.disp.lsfcli.stubCommand = lsfstub{
 233                 errorRate: 0.1,
 234                 sudoUser:  s.disp.Cluster.Containers.LSF.BsubSudoUser,
 235         }.stubCommand(s, c)
 236         s.disp.Start()
 237
 238         deadline := time.Now().Add(20 * time.Second)
 239         for range time.NewTicker(time.Second).C {
 240                 if time.Now().After(deadline) {
 241                         c.Error("timed out")
 242                         break
 243                 }
 244                 // "queuedcontainer" should be running
 245                 if _, ok := s.disp.lsfqueue.Lookup(arvadostest.QueuedContainerUUID); !ok {
 246                         continue
 247                 }
 248                 // "lockedcontainer" should be cancelled because it
 249                 // has priority 0 (no matching container requests)
 250                 if _, ok := s.disp.lsfqueue.Lookup(arvadostest.LockedContainerUUID); ok {
 251                         continue
 252                 }
 253                 // "crTooBig" should be cancelled because lsf stub
 254                 // reports there is no suitable instance type
 255                 if _, ok := s.disp.lsfqueue.Lookup(s.crTooBig.ContainerUUID); ok {
 256                         continue
 257                 }
 258                 var ctr arvados.Container
 259                 if err := s.disp.arvDispatcher.Arv.Get("containers", arvadostest.LockedContainerUUID, nil, &ctr); err != nil {
 260                         c.Logf("error getting container state for %s: %s", arvadostest.LockedContainerUUID, err)
 261                         continue
 262                 } else if ctr.State != arvados.ContainerStateQueued {
 263                         c.Logf("LockedContainer is not in the LSF queue but its arvados record has not been updated to state==Queued (state is %q)", ctr.State)
 264                         continue
 265                 }
 266
 267                 if err := s.disp.arvDispatcher.Arv.Get("containers", s.crTooBig.ContainerUUID, nil, &ctr); err != nil {
 268                         c.Logf("error getting container state for %s: %s", s.crTooBig.ContainerUUID, err)
 269                         continue
 270                 } else if ctr.State != arvados.ContainerStateCancelled {
 271                         c.Logf("container %s is not in the LSF queue but its arvados record has not been updated to state==Cancelled (state is %q)", s.crTooBig.ContainerUUID, ctr.State)
 272                         continue
 273                 } else {
 274                         c.Check(ctr.RuntimeStatus["error"], check.Equals, "There are no suitable hosts for the job;")
 275                 }
 276                 c.Log("reached desired state")
 277                 break
 278         }
 279 }