/// <summary>
/// Returns the number of tasks that failed for reasons other than preemption
/// </summary>
/// <param name="schedulerJob"></param>
/// <returns></returns>
private int GetNonPreemptedFailedTaskCount(IScheduler scheduler, ISchedulerJob schedulerJob)
{
int ret = 0;
try
{
// Filter by failed tasks that failed due to preemption
IFilterCollection fc = scheduler.CreateFilterCollection();
fc.Add(FilterOperator.NotEqual, TaskPropertyIds.FailureReason, FailureReason.Preempted);
fc.Add(FilterOperator.Equal, TaskPropertyIds.State, TaskState.Failed);
// Only return the task Ids
IPropertyIdCollection propIds = new PropertyIdCollection();
propIds.AddPropertyId(TaskPropertyIds.TaskId);
using (ISchedulerRowSet failedTasks = schedulerJob.OpenTaskRowSet(propIds, fc, null, true))
{
ret = failedTasks.GetCount();
}
}
catch (Exception ex)
{
TraceHelper.TraceEvent(this.sessionid, TraceEventType.Warning, "[JobMonitorEntry] Failed to get non-preempted failed task count : {0}", ex);
}
return(ret);
}
/// <summary>
/// Gets the task info
/// </summary>
/// <returns>returns the task info as a dictionary, keyed by task id</returns>
/// <remarks>
/// This method returns a list of task info which ChangeTime property is in this rank: [this.lastChangeTime, DateTime.Now].
/// This method does not change this.lastChangeTime to DateTime.Now after getting tasks because it may fail when sending those information to broker
/// So changeTime is outputed and this.lastChangeTime should be modified to this time after suceeded sending back task info
/// </remarks>
private List <TaskInfo> GetTaskInfo()
{
DateTime changeTime = DateTime.UtcNow;
try
{
TraceHelper.TraceEvent(this.sessionid, TraceEventType.Verbose, "[JobMonitorEntry] Query task info...");
// Step 1: Query task allocation history to fetch node id and core id for tasks
Dictionary <int, TaskAllocationHistoryItem> taskInfoDic = new Dictionary <int, TaskAllocationHistoryItem>();
PropertyIdCollection allocationPropertyCollection = new PropertyIdCollection();
allocationPropertyCollection.AddPropertyId(AllocationProperties.TaskId);
allocationPropertyCollection.AddPropertyId(AllocationProperties.CoreId);
allocationPropertyCollection.AddPropertyId(AllocationProperties.NodeName);
using (ISchedulerRowEnumerator rows = this.schedulerJob.OpenTaskAllocationHistoryEnumerator(allocationPropertyCollection))
{
foreach (PropertyRow row in rows)
{
// Note: Finished/Failed/Canceled task will also be enumerated here
// We are going to add them into the dic regaredless of the state
// because only running tasks will be queried in the following logic.
int objectId = (int)row[AllocationProperties.TaskId].Value;
TaskAllocationHistoryItem taskInfo;
if (taskInfoDic.TryGetValue(objectId, out taskInfo))
{
// For each task instance cache the assigned resource with the lowest coreId. This is needed when node or socket allocation is used
// in order to generate the correct port to connect to the service host
int coreId = (int)row[AllocationProperties.CoreId].Value;
if (taskInfo.FirstCoreId > coreId)
{
taskInfo.FirstCoreId = coreId;
}
taskInfo.Capacity++;
}
else
{
taskInfo = new TaskAllocationHistoryItem();
taskInfo.Capacity = 1;
taskInfo.FirstCoreId = (int)row[AllocationProperties.CoreId].Value;
taskInfo.NodeName = (string)row[AllocationProperties.NodeName].Value;
taskInfoDic.Add(objectId, taskInfo);
}
}
}
TraceHelper.TraceEvent(this.sessionid, TraceEventType.Verbose, "[JobMonitorEntry] Query task info (got task allocation history).");
// Step 2: Get task states from scheduler
// Only task whose change time is between lastChangeTime and Now will be queried
// Only task id and state are required, will get node name according to node id from allocation history got from step 1.
IPropertyIdCollection collection = new PropertyIdCollection();
collection.AddPropertyId(TaskPropertyIds.Id);
collection.AddPropertyId(TaskPropertyIds.State);
FilterCollection fc = new FilterCollection();
fc.Add(FilterOperator.GreaterThan, TaskPropertyIds.ChangeTime, this.lastChangeTime);
fc.Add(FilterOperator.LessThanOrEqual, TaskPropertyIds.ChangeTime, changeTime);
// FIXME: There's performance impact on this query because we look for TaskPropertyIds.Type
// which is requires a table join. Need to have a better way to do so.
fc.Add(FilterOperator.Equal, TaskPropertyIds.Type, TaskType.Service);
List <PropertyRow> taskRows = new List <PropertyRow>();
// The ISchedulerRowSet object is a snapshot and is always a new object, so no lock is needed
foreach (var taskRow in this.schedulerJob.OpenTaskEnumerator(collection, fc, null, true))
{
taskRows.Add(taskRow);
}
TraceHelper.TraceEvent(this.sessionid, TraceEventType.Verbose,
"[JobMonitorEntry] GetTaskInfo, got {0} rows.", taskRows.Count);
if (taskRows.Count == 0)
{
// no service task dispathed yet.
TraceHelper.TraceEvent(this.sessionid, TraceEventType.Warning,
"[JobMonitorEntry] Failed to get task property rows.");
return(null);
}
TraceHelper.TraceEvent(this.sessionid, TraceEventType.Verbose, "[JobMonitorEntry] Query task info (got task info rows from scheduler).");
this.schedulerJob.Refresh();
int jobRequeueCount = this.schedulerJob.RequeueCount;
TraceHelper.TraceEvent(this.sessionid, TraceEventType.Verbose, "[JobMonitorEntry] Job requeue count is {0}", jobRequeueCount);
List <TaskInfo> results = new List <TaskInfo>(taskRows.Count);
foreach (PropertyRow row in taskRows)
{
int objectId = (int)row[TaskPropertyIds.Id].Value;
TaskAllocationHistoryItem taskInfo;
if (!taskInfoDic.TryGetValue(objectId, out taskInfo))
{
continue;
}
TaskState state = (TaskState)row[TaskPropertyIds.State].Value;
if (state == TaskState.Running || state == TaskState.Dispatching)
{
TraceHelper.TraceEvent(this.sessionid, TraceEventType.Information, "[JobMonitorEntry] Task {0} changed into Running", objectId);
string machineName = taskInfo.NodeName;
NodeLocation location = NodeLocation.OnPremise;
string azureServiceName = null;
string azureLoadBalancerAddress = null;
try
{
this.GetNodeInfo(machineName, out location, out azureServiceName, out azureLoadBalancerAddress);
}
catch (Exception e)
{
// if exception happens when querying node info, just skip this node temporarily.
TraceHelper.TraceEvent(this.sessionid, TraceEventType.Warning, "[JobMonitorEntry] -> Get node info for task {0} throws exception. Exception: {1}", objectId, e);
continue;
}
TraceHelper.TraceEvent(this.sessionid, TraceEventType.Verbose, "[JobMonitorEntry] ->Get machine name for task {0}: {1}", objectId, machineName);
int capacity = taskInfo.Capacity;
int coreindex = taskInfo.FirstCoreId;
TraceHelper.TraceEvent(this.sessionid, TraceEventType.Verbose, "[JobMonitorEntry] ->Get coreid for task {0}: {1}", objectId, coreindex);
TraceHelper.TraceEvent(this.sessionid, TraceEventType.Verbose, "[JobMonitorEntry] ->Get AzureLoadBalancerAddress for task {0}: {1}", objectId, azureLoadBalancerAddress);
TaskInfo info = new TaskInfo();
info.Id = objectId;
info.Capacity = capacity;
if (SoaHelper.IsOnAzure())
{
info.MachineVirtualName = machineName;
}
else
{
info.MachineName = machineName;
}
info.Location = NodeLocationConverter.FromHpcNodeLocation(location);
info.ProxyServiceName = azureServiceName;
info.AzureLoadBalancerAddress = azureLoadBalancerAddress;
info.State = TaskStateConverter.FromHpcTaskState(state);
info.FirstCoreIndex = coreindex;
info.JobRequeueCount = jobRequeueCount;
results.Add(info);
}
else if (state == TaskState.Failed || state == TaskState.Canceled || state == TaskState.Canceling || state == TaskState.Finished || state == TaskState.Finishing)
{
TaskInfo info = new TaskInfo();
info.Id = objectId;
info.State = TaskStateConverter.FromHpcTaskState(state);
info.JobRequeueCount = jobRequeueCount;
results.Add(info);
}
}
this.lastChangeTime = changeTime;
return(results);
}
catch (Exception ex)
{
TraceHelper.TraceEvent(this.sessionid, TraceEventType.Warning, "[JobMonitorEntry] Fail when get task info: {0}", ex);
return(null);
}
finally
{
TraceHelper.TraceEvent(this.sessionid, TraceEventType.Verbose, "[JobMonitorEntry] Query task info finished.");
}
}
/// <summary>
/// Calculate the min and max value for the service job
/// </summary>
private void CalculateMinAndMax()
{
PropertyId[] propIds;
PropertyRow row;
int userMax, userMin;
switch (this.schedulerJob.UnitType)
{
case JobUnitType.Node:
userMax = this.schedulerJob.MaximumNumberOfNodes;
userMin = this.schedulerJob.MinimumNumberOfNodes;
propIds = new PropertyId[] { JobPropertyIds.ComputedMaxNodes, JobPropertyIds.ComputedMinNodes };
break;
case JobUnitType.Socket:
userMax = this.schedulerJob.MaximumNumberOfSockets;
userMin = this.schedulerJob.MinimumNumberOfSockets;
propIds = new PropertyId[] { JobPropertyIds.ComputedMaxSockets, JobPropertyIds.ComputedMinSockets };
break;
default:
userMax = this.schedulerJob.MaximumNumberOfCores;
userMin = this.schedulerJob.MinimumNumberOfCores;
propIds = new PropertyId[] { JobPropertyIds.ComputedMaxCores, JobPropertyIds.ComputedMinCores };
break;
}
IFilterCollection filter = new FilterCollection();
filter.Add(FilterOperator.Equal, JobPropertyIds.Id, this.schedulerJob.Id);
IPropertyIdCollection property = new PropertyIdCollection();
foreach (PropertyId pid in propIds)
{
property.AddPropertyId(pid);
}
using (ISchedulerRowSet set = this.scheduler.OpenJobRowSet(property, filter, null))
{
PropertyRow[] rows = set.GetRows(0, set.GetCount() - 1).Rows;
Debug.Assert(rows.Length > 0);
row = rows[0];
}
string callerName = "[JobMonitorEntry.GetMinAndMax]";
int computedMax = (int)JobHelper.GetStorePropertyValue(row.Props[0], propIds[0], 0, callerName);
int computedMin = (int)JobHelper.GetStorePropertyValue(row.Props[1], propIds[1], 0, callerName);
if (this.schedulerJob.CanShrink)
{
this.minUnits = this.schedulerJob.AutoCalculateMin ? computedMin : userMin;
}
else
{
this.minUnits = userMin;
}
if (this.schedulerJob.CanGrow)
{
this.maxUnits = this.schedulerJob.AutoCalculateMax ? computedMax : userMax;
}
else
{
this.maxUnits = userMax;
}
if (this.maxUnits < this.minUnits)
{
this.maxUnits = this.minUnits;
}
TraceHelper.TraceEvent(this.sessionid, TraceEventType.Information, "[JobMonitorEntry] MaxUnits = {0}, MinUnits = {1}", this.maxUnits, this.minUnits);
}