Steps to performance tuning

Notes from Practical Approaches to Great App Performance

• measure existing performance, set up baseline.
• re-measure performance, compare, document, share.
• focus on total impact (vs. just an area that are not used by many users, not used frequently)
• reproduce -> profile -> modify -> repeat
• good to have automated performance tests to avoid regressions over time (the thousand performance paper cuts)
• important to understand user usage pattern, so that target is clear and optimizing on the most valuable areas.
• break large task into smaller steps and write unit tests for each step. That helps pinpoint issue.
• integration tests are measured from UX perspective, covering not just the task, but all other areas that work together.
• always start with integration tests, so that you know what area to start the tuning.
• use Time Profiler to get performance profile.
• if not debugging threads related, better not to group the trace by thread. To disable it: Call Tree -> uncheck Separate by Thread.
• how to remove noise, or focus on the signal
  • focus on one thread at a time. To do that: select only the thread that cost the most in the Track Filter panel
  • focus on one message at a time. To do that: select the suspecious in the Heaviest back trace panel
  • remove recursions. To do that: Call Tree -> check Flatten recursion.
  • drill down (bring up to the top level). To do that: right click on the trace -> Focus on subtree
  • hide all objc related messages in call tree (bubble those “cost of doing business” up to the parent callers). To do that: right click -> Charge 'libobjc.A.dylib' to callers
  • hide all small “contributors”. Say for a 2-sec sample range, filter out only trace that > 20ms, that means only focusing on contributors that cost >1%. To do that: Call trace constraints -> 20 for min.
• even the slowness happens in system lib, there is possibilities to optimize
  • the data you passed into the system lib to operate
  • how many times you are calling this system method
  • system method is calling back into your code via delegates
• key-value observer (KVO) (i.e., “update UI whenever model changes”) in a loop could impact performance
• instant app launch means? 500ms to 600ms (that’s how long it takes the zoom animation from the home screen)
• do initialization, such as DB init, in the background thread, to improve launch time.
• as little work as possible in the initializers
• load data on screen synchronously, off-screen data async.
• strive for constant time.