Performance: forks, pipelining, async

Where time goes: for each task and host, Ansible opens an SSH session, creates a temp directory, copies the module, runs it and cleans up. Optimisations attack each part:

• forks (default 5): how many hosts run a task in parallel. Raise to 25-100 for big fleets, limited by control-node CPU and memory.
• SSH multiplexing: ControlMaster/ControlPersist (on by default in Ansible's ssh args) reuses one TCP+auth connection for many tasks.
• pipelining (pipelining = True): sends the module over the existing session's stdin instead of copying a file, cutting several SSH operations per task. Requires that sudo doesn't need requiretty.
• Facts: gather_facts: false when not needed, gather_subset to collect less, or fact caching (jsonfile, redis) with gathering = smart to reuse facts between runs.
• Strategies: the default linear waits for all hosts at each task; free lets each host run ahead independently. serial batches trade speed for safety.
• async: async: 600, poll: 0 starts a long task (a big download, a backup) in the background and moves on; check it later with async_status.
• Fewer tasks: pass lists to package modules instead of looping, and use template once rather than many lineinfile edits.

Measure before tuning: the profile_tasks and timer callbacks (callbacks_enabled = ansible.posix.profile_tasks, ansible.posix.timer) show per-task timings.