You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

train.go 7.0 kB

4 months ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225
  1. package tasksync
  2. import (
  3. "errors"
  4. "fmt"
  5. "github.com/zeromicro/go-zero/core/logx"
  6. "gitlink.org.cn/JointCloud/pcm-coordinator/internal/config"
  7. "gitlink.org.cn/JointCloud/pcm-coordinator/internal/scheduler/database"
  8. "gitlink.org.cn/JointCloud/pcm-coordinator/internal/scheduler/service/collector"
  9. "gitlink.org.cn/JointCloud/pcm-coordinator/internal/scheduler/service/utils/jcs"
  10. "gitlink.org.cn/JointCloud/pcm-coordinator/internal/types"
  11. "gitlink.org.cn/JointCloud/pcm-coordinator/pkg/constants"
  12. "gitlink.org.cn/JointCloud/pcm-coordinator/pkg/models"
  13. "google.golang.org/grpc/codes"
  14. "google.golang.org/grpc/status"
  15. "net/http"
  16. "strconv"
  17. "sync"
  18. )
  19. type SyncTrain struct {
  20. mu sync.Mutex
  21. aiStorages *database.AiStorage
  22. aiCollectorAdapterMap map[string]map[string]collector.AiCollector
  23. config *config.Config
  24. }
  25. func NewTrainTask(storage *database.AiStorage, aiCollectorAdapterMap map[string]map[string]collector.AiCollector, config *config.Config) *SyncTrain {
  26. return &SyncTrain{
  27. aiStorages: storage,
  28. aiCollectorAdapterMap: aiCollectorAdapterMap,
  29. config: config,
  30. }
  31. }
  32. func (s *SyncTrain) UpdateAiTaskStatus(tasklist []*types.TaskModel) {
  33. s.mu.Lock()
  34. defer s.mu.Unlock()
  35. list := make([]*types.TaskModel, len(tasklist))
  36. copy(list, tasklist)
  37. for i := len(list) - 1; i >= 0; i-- {
  38. if list[i].AdapterTypeDict != "1" || list[i].Status == constants.Succeeded || list[i].Status == constants.Failed {
  39. list = append(list[:i], list[i+1:]...)
  40. }
  41. }
  42. if len(list) == 0 {
  43. return
  44. }
  45. for _, task := range list {
  46. aiTaskList, err := s.aiStorages.GetAiTaskListById(task.Id)
  47. if err != nil {
  48. logx.Errorf("UpdateAiTaskStatus Get AiTask Error %s", err.Error())
  49. }
  50. if len(aiTaskList) == 0 {
  51. continue
  52. }
  53. go s.updateAiTask(aiTaskList)
  54. }
  55. }
  56. func (s *SyncTrain) UpdateTaskStatus(tasklist []*types.TaskModel) {
  57. s.mu.Lock()
  58. defer s.mu.Unlock()
  59. list := make([]*types.TaskModel, len(tasklist))
  60. copy(list, tasklist)
  61. for i := len(list) - 1; i >= 0; i-- {
  62. if list[i].AdapterTypeDict != "1" || list[i].Status == constants.Succeeded || list[i].Status == constants.Failed || list[i].Status == constants.Cancelled {
  63. list = append(list[:i], list[i+1:]...)
  64. }
  65. }
  66. if len(list) == 0 {
  67. return
  68. }
  69. for _, task := range list {
  70. aiTask, err := s.aiStorages.GetAiTaskListById(task.Id)
  71. if err != nil {
  72. logx.Errorf("UpdateTaskStatus Get AiTask Error %s", err.Error())
  73. }
  74. if len(aiTask) == 0 {
  75. continue
  76. }
  77. logx.Errorf("############ Report Status Message Before switch %s", task.Status)
  78. if len(aiTask) == 1 {
  79. logx.Errorf("############ Report Status Message Switch %s", aiTask[0].Status)
  80. switch aiTask[0].Status {
  81. case constants.Completed:
  82. task.Status = constants.Succeeded
  83. logx.Errorf("############ Report Status Message Before Sending %s", task.Status)
  84. err = s.reportStatusMessages(task, aiTask[0], true)
  85. if err != nil {
  86. logx.Errorf("reportStatusMessages Error %s", err.Error())
  87. }
  88. case constants.Failed:
  89. task.Status = constants.Failed
  90. logx.Errorf("############ Report Status Message Before Sending %s", task.Status)
  91. err = s.reportStatusMessages(task, aiTask[0], false)
  92. if err != nil {
  93. logx.Errorf("reportStatusMessages Error %s", err.Error())
  94. }
  95. default:
  96. task.Status = aiTask[0].Status
  97. }
  98. task.StartTime = aiTask[0].StartTime
  99. task.EndTime = aiTask[0].EndTime
  100. err := s.aiStorages.UpdateTask(task)
  101. if err != nil {
  102. logx.Errorf("UpdateTaskStatus Update Task Error %s", err.Error())
  103. }
  104. }
  105. }
  106. }
  107. func (s *SyncTrain) updateAiTask(aiTaskList []*models.TaskAi) {
  108. var wg sync.WaitGroup
  109. for _, aitask := range aiTaskList {
  110. t := aitask
  111. if t.Status == constants.Completed || t.Status == constants.Failed || t.JobId == "" || t.Status == constants.Cancelled {
  112. continue
  113. }
  114. wg.Add(1)
  115. go func() {
  116. h := http.Request{}
  117. trainingTask, err := s.aiCollectorAdapterMap[strconv.FormatInt(t.AdapterId, 10)][strconv.FormatInt(t.ClusterId, 10)].GetTrainingTask(h.Context(), t.JobId)
  118. if err != nil {
  119. if status.Code(err) == codes.DeadlineExceeded {
  120. msg := fmt.Sprintf("###UpdateAiTaskStatus###, AiTaskId: %v, clusterId: %v , JobId: %v, error: %v \n", t.Id, t.ClusterId, t.JobId, err.Error())
  121. logx.Errorf(errors.New(msg).Error())
  122. wg.Done()
  123. return
  124. }
  125. msg := fmt.Sprintf("###UpdateAiTaskStatus###, AiTaskId: %v, clusterId: %v , JobId: %v, error: %v \n", t.Id, t.ClusterId, t.JobId, err.Error())
  126. logx.Errorf(errors.New(msg).Error())
  127. wg.Done()
  128. return
  129. }
  130. if trainingTask == nil {
  131. wg.Done()
  132. return
  133. }
  134. switch trainingTask.Status {
  135. case constants.Running:
  136. if t.Status != trainingTask.Status {
  137. s.aiStorages.AddNoticeInfo(strconv.FormatInt(t.AdapterId, 10), t.AdapterName, strconv.FormatInt(t.ClusterId, 10), t.ClusterName, t.Name, "running", "任务运行中")
  138. t.Status = trainingTask.Status
  139. }
  140. case constants.Failed:
  141. if t.Status != trainingTask.Status {
  142. s.aiStorages.AddNoticeInfo(strconv.FormatInt(t.AdapterId, 10), t.AdapterName, strconv.FormatInt(t.ClusterId, 10), t.ClusterName, t.Name, "failed", "任务失败")
  143. t.Status = trainingTask.Status
  144. }
  145. case constants.Completed:
  146. if t.Status != trainingTask.Status {
  147. s.aiStorages.AddNoticeInfo(strconv.FormatInt(t.AdapterId, 10), t.AdapterName, strconv.FormatInt(t.ClusterId, 10), t.ClusterName, t.Name, "completed", "任务完成")
  148. t.Status = trainingTask.Status
  149. }
  150. default:
  151. if t.Status != trainingTask.Status {
  152. s.aiStorages.AddNoticeInfo(strconv.FormatInt(t.AdapterId, 10), t.AdapterName, strconv.FormatInt(t.ClusterId, 10), t.ClusterName, t.Name, "pending", "任务pending")
  153. t.Status = trainingTask.Status
  154. }
  155. }
  156. t.StartTime = trainingTask.Start
  157. t.EndTime = trainingTask.End
  158. err = s.aiStorages.UpdateAiTask(t)
  159. if err != nil {
  160. msg := fmt.Sprintf("###UpdateAiTaskStatus###, AiTaskId: %v, clusterId: %v , JobId: %v, error: %v \n", t.Id, t.ClusterId, t.JobId, err.Error())
  161. logx.Errorf(errors.New(msg).Error())
  162. wg.Done()
  163. return
  164. }
  165. wg.Done()
  166. }()
  167. }
  168. wg.Wait()
  169. }
  170. func (s *SyncTrain) reportStatusMessages(task *types.TaskModel, aiTask *models.TaskAi, status bool) error {
  171. report := &jcs.JobStatusReportReq{}
  172. reportMsg := &jcs.TrainReportMessage{
  173. Type: "Train",
  174. TaskName: task.Name,
  175. TaskID: strconv.FormatInt(task.Id, 10),
  176. }
  177. var output string
  178. switch aiTask.ClusterName {
  179. case "openI":
  180. output = aiTask.JobId
  181. case "鹏城云脑II-modelarts":
  182. output = aiTask.Output
  183. }
  184. reportMsg.Status = status
  185. reportMsg.Message = ""
  186. reportMsg.ClusterID = strconv.FormatInt(aiTask.ClusterId, 10)
  187. reportMsg.Output = output
  188. report.Report = reportMsg
  189. err := jcs.StatusReport(s.config.JcsMiddleware.JobStatusReportUrl, report)
  190. if err != nil {
  191. return err
  192. }
  193. err = jcs.TempSaveReportToTask(s.aiStorages, task, report)
  194. if err != nil {
  195. return err
  196. }
  197. return nil
  198. }

PCM is positioned as Software stack over Cloud, aiming to build the standards and ecology of heterogeneous cloud collaboration for JCC in a non intrusive and autonomous peer-to-peer manner.