Fine-tuning PE placement assignments for an instance

To change the placement of a processing element (PE), you must restart it. Teracloud® Streams provides recommendations that you can use to fine-tune the PE placements. In general, the goal of the recommendations is to move PEs from resources that are overloaded to a new resource that is underloaded. You can use the Streams Console or streamtool commands to restart PEs with the recommendations. This procedure uses the Streams Console.

About this task

After a job runs for a while, you might want to adjust the placement of some PEs due to load imbalances across the resources in your Streams instance. For example, if you notice that one resource is more heavily loaded than the other resources, you might want to recompute the placement of the PEs on the overloaded resource so that the load is more balanced across all the resources.
Note: New PE placement must meet the constraints as specified in the job and better placement is not always feasible.

Because the instance continues to run during the analysis and the recommendations are computed from a snapshot of data, the scheduler does not return detailed recommendations about which candidate resource should be the target for the move. If the data changed since the recommendation, the scheduler might choose a different resource when the PE is restarted.

The upper and lower threshold settings for these three metrics play a key part in the recommendation analysis:

  • CPU utilization
  • Memory utilization
  • Network bandwidth utilization

If the value for any of these metrics for a resource exceeds its upper threshold, the resource is classified as overloaded. If the values for all three metrics for a resource go below their lower thresholds, the resource is classified as underloaded.

To modify threshold values, see Changing resource usage thresholds.

Procedure

  1. Determine the scope of your analysis. Scope choices are for a job, an instance, or for resources. The scheduler computes only PE restart recommendations for PEs that are part of jobs for which the requester has the proper authority.
  2. In the Streams Console, go to the Streams Tree card in the Application dashboard. Click the Properties information icon for an instance or for a job, and then click Show PE Restart Recommendations. You can select specific resources in the dialog that follows after you click Show PE Restart Recommendations for an instance.

    After the analysis begins, the Streams job scheduler returns one recommendation for every PE. The strength of the recommendation depends on the recommendation category.

  3. Select a recommendation category:
    Preferred
    The best candidate PEs to be restarted on other hosts in the instance. This is the strongest recommendation. There might be more than one recommendation in this category. In some cases, there might not be a recommendation that qualifies as Preferred.
    High
    The current resource running the PE is overloaded and there is at least one candidate resource that is underloaded. This is a strong recommendation. There can be many recommendations in this category.
    Directed
    The current resource is overloaded, but there are no alternate resources that are underloaded. This recommendation might be a good choice if you add new resources.
    Low
    There are some candidate resources that have lighter loads than the current resource. If you select this recommendation, it might improve the results. However, the improvement is probably not as pronounced as those resulting from selecting a Preferred or High recommendation.
    None
    Do not restart the PE now. The PE is not movable or there is not an alternate resource that is less loaded.
  4. To invoke a recommendation, select a PE restart recommendation entry in the table, and then click Restart Chosen PEs.
    You can invoke multiple recommendations, but each recommendation is made independently. Invoking multiple recommendations together might or might not provide better results. To confirm how much the PE restarts have helped, allow the new placements to run for a while with a representative amount of the workload, and then request a new set of restart recommendations. Click Show PE Restart Recommendations to update the PE Restart Recommendations list.