[ https://issues.apache.org/jira/browse/HELIX-655?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15993887#comment-15993887 ]
Jiajun Wang commented on HELIX-655: ----------------------------------- h1. Design h2. Task Throttling Per Participant This limitation conceptually equals to max thread pool size in TaskStateModelFactory on the participant. If user constructs TaskStateModelFactory using their customized executor with a limited sized thread pool, that participant will never executing more tasks than the threshold. The problem is that since the limitation is not known by the controller, tasks will still be assigned to the participant. And they will be queued in participant thread pool and never re-assigned. It makes more sense to throttle tasks in the controller. At the same time that tasks are assigned to participants. Basically, participant is configured with a "MaxRunnigTasksNumber". And the controller assigns task accordingly. h3. pseudo code When calculating Best possible state in the JobRebalancer Foreach Job in RunnableJobs: TaskToParticipantMapping = CalculateAssignment(Job) Foreach MappingEntry in TaskToParticipantMapping: If Running_task + ToBeAssigned_task exceeds Participant_Task_Threshold: TaskToParticipantMapping.remove(MappingEntry) Try next applicable participant (consider task attached to resource) The above logic can be considered as a task queue algorithm. However, the original assignment will keep relying on current logic. So if all participants have enough capacity, tasks will still be evenly dispatched. h3. participant configuration { "id" : "localhost_12918", "simpleFields" : { "HELIX_ENABLED" : "true", "HELIX_ENABLED_TIMESTAMP" : "1493326930182", "HELIX_HOST" : "localhost", "HELIX_PORT" : "12918", "MAX_RUNNING_TASK" : "55" } } h3. Backward compatible For old participants, controller assumes the thread pool is with a default capacity 40 (equal to default message handling thread pool size). h4. Assumption Note that if some tasks have workload that are much heavier than others, only control tasks number won't work. In this design, we assume that tasks have the approximately same workload. h3. [Stretch] Performance optimization Existing JobRebalancer will be trigger everytime a state change event happens. That means completely sorting all pending jobs/tasks and calculate assignment. A better strategy is maintain a Job priority queue in the controller. When a job became runnable, enqueue. When a job is complete, dequeue. Any task state update, check paticipant capacity and assign task from the queue if possible. This refactoring is considered as a stretch goal. h3. Alternative option h4. "GlobalMaxConcurrentJobNumber" Per Cluster Helix controller restricts the number of running jobs. However, with this throttling, once a job is scheduled, it will occupy the slot until finish. This will be bad when all the running jobs are long-run. No new jobs will be scheduled. Moreover, it's harder for admin to set a reasonable total job count, given workflows and jobs are usually quite different regarding their real workload. Comparing these 2 options, "MaxRunnigTasksPerParticipant" is directly related to participant's capacity. Once the controller schedule tasks according to this, we can for sure avoid overloading the instances. Even we throttle jobs, there is no guarantee about the running thread in each participant. Moreover, user can currently control job scheduling by adjusting the frequency of submitting jobs. So "GlobalMaxConcurrentJobNumber" is not necessary. h2. Job Priority Given limited resource, which job we schedule first? h3. Schedule the jobs with the highest priority first until participants are full In this design, we proposed the simplest solution for priority control. User can configure job resource priority or Helix will assume "age" (time that the job was scheduled) as priority. If part of the jobs are assigned priority, others are not, Helix will assume jobs with priority setting have higher priority. One issue here is that if the application keeps sending high priority jobs to Helix, lower priority jobs will be starving. Since this is controlled by the application (and mostly desired result), Helix won't apply any additional control on these starving jobs. h3. Alternative options h4. Option 1. Using per-job and per-workflow concurrency control to implement priority WorkflowConfig.ParallelJobs and JobConfig.numConcurrentTasksPerInstance are used to control how many jobs and tasks can be executed in parallel within a single workflow. Given that the cluster administrators can configure these numbers "correctly", workflows will be assigned expected resources eventually. However, there is no promising that high priority workflows will be scheduled before others. This is because tasks are picked up randomly, so the controller may end up with fill the task pool with all items from low priority workflows. Cons # Hard for users to setup the right numbers. # Cannot strictly ensure priority. # May lead to low utilization. h4. Option 2. Jobs are assigned with execution slots according to priority Helix controller assigns different portions of resources (execute slots) to jobs according to their priority. For instance, we may have following assignment given the total capacity is 100. So high priority jobs will always get a larger portion of resources. If any job does not use all of its portions, our algorithm should be smart enough to assign those portion to other jobs. The problem of this method is complixity. In addition, since we are not assigning all possible resource to the highest priority jobs, those jobs are not guaranteed to be finished quickly, and users might feel confusing. > Helix per-participant concurrent task throttling > ------------------------------------------------ > > Key: HELIX-655 > URL: https://issues.apache.org/jira/browse/HELIX-655 > Project: Apache Helix > Issue Type: New Feature > Components: helix-core > Affects Versions: 0.6.x > Reporter: Jiajun Wang > > h1. Overview > Currently, all runnable jobs/tasks in Helix are equally treated. They are all > scheduled according to the rebalancer algorithm. Means, their assignment > might be different, but they will all be in RUNNING state. > This may cause an issue if there are too many concurrently runnable jobs. > When Helix controller starts all these jobs, the instances may be overload as > they are assigning resources and executing all the tasks. As a result, the > jobs won't be able to finish in a reasonable time window. > The issue is even more critical to long run jobs. According to our meeting > with Gobblin team, when a job is scheduled, they allocate resource for the > job. So in the situation described above, more and more resources will be > reserved for the pending jobs. The cluster will soon be exhausted. > For solving the problem, an application needs to schedule jobs in a > relatively low frequency (what Gobblin is doing now). This may cause low > utilization. > A better way to fix this issue, at framework level, is throttling jobs/tasks > that are running concurrently, and allowing setting priority for different > jobs to control total execute time. > So given same amount of jobs, the cluster is in a better condition. As a > result, jobs running in that cluster have a more controllable execute time. > Existing related control mechanisms are: > * ConcurrentTasksPerInstance for each job > * ParallelJobs for each workflow > * Threadpool limitation on the participant if user customizes > TaskStateModelFactory. > But none of them can directly help when concurrent workflows or jobs number > is large. If an application keeps scheduling jobs/jobQueues, Helix will start > any runnable jobs without considering the workload on the participants. > The application may be able to carefully configures these items to achieve > the goal. But they won't be able to easily find the sweet spot. Especially > the cluster might be changing (scale out etc.). > h2. Problem summary > # All runnable tasks will start executing, which may overload the participant. > # Application needs a mechanism to prioritize important jobs (or workflows). > Otherwise, important tasks may be blocked by other less important ones. And > allocated resource is wasted. > h3. Feature proposed > Based on our discussing, we proposed 2 features that can help to resolve the > issue. > # Running task throttling on each participant. This is for avoiding overload. > # Job priority control that ensures high priority jobs are scheduled earlier. > In addition, application can leverage workflow/job monitor items as feedback > from Helix to adjust their stretgy. -- This message was sent by Atlassian JIRA (v6.3.15#6346)