Optimizing Apache Storm Topologies

1 minute read

This article summarizes hints for optimizing and deploying Apache Storm topologies.

Setup your storm cluster

  1. I/O is zookeeper’s main bottleneck - ensure that the /data partition of zookeeper machines serializes to quick storage (ramdisk ;)
  2. Determine the number of parallelism units using the following rule of thumb:
    • number of available CPU cores on all machines minus one core per machine that is used for the Acker
    • Example: 2 machines with 48 and 1 machine with 32 cores; parallelism units = 2x(48-1) + (32-1) = 125
  3. Using multiple workers per machine allows deploying multiple topologies at once (the number of workers is determined by the number of ports configured in the supervisor.slots.ports setting in storm.yaml)

Topology configuration suggestions

  1. Use one worker per machine and topology (intra-worker transports are more efficient)
  2. The number of executors depends on whether your bolt is I/O or CPU bound
    • CPU bound: configure one executor per available parallelism unit
    • I/O bound: use 10-100 workers per parallelism unit, depending on the expected I/O delay
  3. The total number of parallelism units in your topology should equal the number of available parallelism units

Profiling the topology

  1. Storm UI: use the capacity metric to identify bolts which require a higher parallelism
  2. your nextTuple and execute methods determine the spout’s/bolt’s runtime - optimize these methods
  3. use queue’s for I/O in spouts or terminal bolts (i.e. write final results to a queue and use a writer thread that performs batch inserts to serialize the queue to disk)


  • workers process - responsible for executing the topology on a particular machine
  • executor - thread spawned by the worker for a particular component (bold or spout); the number of executors is configured by setting the parallelism hint parameter in the setSpout or setBolt method.
  • task - number of instances of a particular bolt/spout to deploy; configuring more than one task using setNumTasks(n) allows to later increase the number of executors for that particular spout/bolt without redeploying the topology.


  1. Scaling Apache Storm
  2. Understanding the Parallelism of an Apache Storm Topology
  3. Stack overflow - What is a “Task” in Storm Parallelism
  4. Hortonworks - Storm Parallelism

Leave a comment