Performance Tuning iOS – Making Mobile Fast

| October 26, 2012

Disclaimer: The following post is intended for a somewhat technical audience. If you come across any unfamiliar words or phrases, please refer to your local software developer.

The problem

Our iOS app has suffered from pretty abysmal performance issues for any groups with more than a couple dozen users and rooms. Since joining Atlassian, we’ve experienced the issue first hand even more acutely (the Atlassian HipChat group has hundreds of users and rooms). We had to do something about it.

The analysis

Before any good performance update comes the part where you close your eyes and get the raw numbers about how slow your current app is. Some of our findings:

  1. Avoid blocking HTTP requests. Before we could even try to connect to our chat servers, we had to hit an HTTP API to get user and server information. This request alone would take ~1.5 seconds to complete on WiFi and a whopping 3s on average for a 3G connection.
  2. Minimize round trips to the server. XMPP has a well-defined spec for securing the chat connection. Unfortunately, it involves no less than 3 back-and-forths between the client and server. Add authentication and session setup onto that and you end up with a good 6 back and forths.
  3. Avoid updating the UI when getting lots of data. Apart from basic network latency, the other thing that makes the app slow is doing the actual work of reading XMPP and setting up all the stuff you actually see. This turned out to be costing a ton of resources since each time we received a user’s presence (the data that tells us whether they’re available/away/dnd, etc), we would update the Lobby (assuming you were viewing the Lobby, which was quite likely since that’s where the app opens to). Calculated out, it turned out that we were spending 12-15 seconds just handling everyone’s presence for a big group like Atlassian. Definitely unacceptable.
  4. Stringprep –  Y U SO SLOW. Most XMPP libraries use a process called “stringprep” when creating JID objects (JIDs are the unique user identifying strings used in XMPP). Stringprep ensures that your JIDs are valid and conform to the spec. Unfortunately, we found that stringprep was responsible for nearly 50% of our CPU usage on the phone (we create lots of JIDs). Fortunately, we run a closed system where we can validate JIDs beforehand. This became the easiest fix of all: remove stringprep 🙂
  5. iOS offers a sophisticated local storage for a reason. Before this update, we made minimal use of local storage on the phone. The problem was that we were just clearing it out completely when you reconnected. Not exactly the best use of the Core Data framework

The changes

  1. Previously on connection, something like this happened (times are for a 3G connection in San Francisco):
    1. PHONE: HTTP Request to get info about which chat server to connect to (~1-3 seconds)
    2. WEB SERVER: “Here’s the information you requested – use it to connect to SERVER”
    3. PHONE: “Hey SERVER, I’m gonna start a session” *(0.3 seconds)
    4. SERVER: “Ok – but first we gotta make sure nobody can snoop on our conversation. Let’s use something called TLS”
    5. PHONE: “Oh, ok, I know TLS! I officially request we secure using TLS” (0.3 seconds)
    6. SERVER: “You may proceed” … <at this point, the TLS handshake happens and all data sent and received is encrypted> (0.4 – 0.6 seconds)
    7. PHONE: “Phew – ok, let’s start that connection again” (0.3 seconds)
    8. SERVER: “Connection started, but first, you need to authenticate”
    9. PHONE: “Got it – here’s the username/password info my user provided” (0.3 seconds)
    10. SERVER: “Glorious success! Your credentials were correct.”
    11. PHONE: “Great, let’s really start a chat session now” (0.3 seconds)
    12. SERVER: “Sure thing – but wait, I need an identifier for this session so I can tell it apart from that connection you made from your desktop”
    13. PHONE: “Oh, ok – how about we call it the ‘iphone’ session?” (0.3 seconds)
    14. SERVER: “That is acceptable – here is your full identifying string…”
    15. PHONE: “Cool, now I officially am starting a session…” <after this we actually start getting chat data> Total time: 3.2s – 5.4s – just to start getting actual chat data from the server
  2. On a 3G connection, when each request to the server can take up to several hundred milliseconds, this is pretty terrible. So we set out to reduce the number of required requests as much as possible. This is what we came up with:
    1. <we already know the server to hit, no need to ask WEB SERVER anymore> <the SSL handshake happens as soon as we open the connection to SERVER> (0.4 seconds) PHONE:  “Hey SERVER, let’s start a session” (0.3 seconds)
    2. SERVER: “Sure thing – connection has been opened. Please choose an auth method: <includes a special HipChat-only auth>”
    3. PHONE: “Here’s the authentication data, including the identifier for my session” (0.3 seconds)
    4. SERVER: “Excellent – authentication validated. Here is your full identifying string. You may now request data” Total time: 1 second
  3. Finally, the other major app slowdown happened when we received the flood of initial presences right after logging in. These triggered a cascading update of displayed data in the Lobby. And as anyone who has done UI performance tuning before, having lots of display drawing to do can cripple a user experience. Our solution: have the server let us know when the flood of presences is done, then update the UI. The result: 80% fewer cycles spent during the connection process.

For the future…

Unfortunately, when we began writing the iOS app, it was on the cusp of the release of iOS 4 (which plagued some devices with performance problems), so we were hesitant to use any of the brand-new features. We still have work to do to move our app over to ARC. We still don’t make nearly enough use of Grand Central Dispatch (mostly because we don’t use the most current version of XMPPFramework). We also have some server improvements to make that will let us further speed up the app for really big teams. XMPP offers a spec for roster versioning which could replace the full roster refreshes we do now. This is just a first step to making the app more usable for larger teams (and part of how we’re making sure that our Native OSX App is going to be lightning fast). The update is available on the App Store today. Download it and let us know what you think!