This guide provides basic steps at tuning a topology to utilize resources efficiently. Currently, resources are primarily measured in terms of CPU cores and RAM. In Heron, some of the basic parameters that are available to tune a topology are, but not limited to, the following:
- Container RAM
- Container CPU
- Component RAMs
- Component Parallelisms
- Number of Containers
Note that tuning a topology may be difficult and may take multiple iterations. Before proceeding, please make sure you understand concepts to understand the terminology, as well as the reasoning behind taking these steps.
Steps to Tune a Topology
Launch the topology with an initial estimate of resources. These can be based on input data size, component logic, or experience from another working topology.
Resolve any backpressure issues by increasing the parallelism or container RAM, or CPU, or appropriately if backpressure is due to an external service.
Make sure there is no spout lag. In steady state, the topology should be able to read the whole of data.
Repeat steps 2 and 3 until there is no backpressure and no spout lag.
By now, the CPU usage and RAM usage are stable. Based on daily of weekly data trends, leave appropriate room for usage spikes, and cut down the rest of the unused resources allocated to topology.
While these steps seem simple, it might take some time to get the topology to its optimal usage. Below are some of the tips that can be helpful during tuning or in general.
If component RAMs for all the components is provided, that will the RAM assigned to those instances. Use this configuration according to their functionality to save of resources. By default, every instance is assigned 1GB of RAM, which can be higher that what it requires. Note that if container RAM is specified, after setting aside some RAM for internal components of Heron, rest of it is equally divided among all the instances present in the container.
A memory intensive operation in bolts can result in GC issues. Be aware of objects that might enter old generation, and cause memory starvation.
You can use
Schemes in spouts to sample down the data. This can helpful when dealing with issues if writing to external services, or just trying to get an early estimate of usage without utilizing much resources. Note that this would still require 100% resource usage in spouts.