Dynamic frameworks load time?

[marcelofabri]

I'm seeing large load times when using dynamic frameworks on my app and decided to check if this was an issue for you guys too.

I could reproduce the issue on an iPad 3 here and sometimes the app won't even launch due to the WatchDog. Am I missing something here? Do you really suffer from this issue too?

This seems to be the same issue as artsy/eigen#586.

[orta]

Are you on iOS9?

[marcelofabri]

iOS 8.1

[orta]

We definitely saw a longer launch time for eigen, it was significant on iOS 9 devices ( after being compiled and deployed from Xcode 6 ) but it wasn't worrying on eigen ( which has way more Pods ) for the last release.

[ashfurrow]

I've seen long launch times, but nothing that would stop the app from launching. Maybe the hardware we run on is fast enough. Iit hasn't been a pressing issue for us on this app since it's an internal tool.

[marcelofabri]

Aug 3 17:05:29 iPad-3-iOS-81 ReportCrash[495] : net.artsy.kiosk.beta failed to launch in time
Aug 3 17:05:29 iPad-3-iOS-81 ReportCrash[495] : Elapsed total CPU time (seconds): 22.320 (user 22.320, system 0.000), 55% CPU

I'm wondering how aren't there more people talking about this problem on their apps.

I mean, we're almost dropping dynamic frameworks because of this. I also opened a DTS with Apple to see if there's any official word on this.

[ashfurrow]

I think @alloy did some testing on eigen and the problem primarily affects debug builds. There's also a helpful command there to shorten the time.

[marcelofabri]

Yeah, I checked that, but I'm also seeing the issue on release builds. Also, if I understood correctly, that command will statically link all frameworks and (I think) Swift frameworks need to be dynamically linked.

[brianmichel]

I've been seeing this recently in our app as well, doing some more investigating and lead me here. You are not alone :)

[orta]

I don't think we have either: Asked on the Apple Dev Forums, or done a DTS support ticket. Both would expand the surface area of people who might be in the know about ways around it.

Other ideas, @neonacho pointed me at https://github.com/johnno1962/Accelerator last night

[marcelofabri]

Here is our DTS response:

Thank you for contacting Apple Developer Technical Support (DTS).

I discovered that these long load times only happen when the debugger is attached. If you select Product > Profile and look at the console in Instruments instead, you will notice the time is in milliseconds instead of seconds.

dyld reports a higher time when the debugger is attached because the debugger halts execution after each framework is loaded in order to gather information. That time that the process is halted counts against the dyld load time because dyld uses mach_absolute_time() to record the time at the start and end of loading frameworks and then prints the difference.

If you are seeing poor framework loading performance without a debugger attached, please file a bug so our dyld engineers can take a look. Combining frameworks may reduce the loading time, but you will hit a lower bound of frameworks since you can’t combine the Swift frameworks bundled with your app.

However, our app still takes 8s to load (7.4s on "total images loading time"). I just responded to the DTS and I'm waiting for a response. On the meantime, I'll test https://github.com/johnno1962/Accelerator.

[alloy]

@marcelofabri Yeah that’s basically the same as reported on the Eigen thread. In our case startup is a bit slower the first time you launch the app, thereafter it goes smoother.

[alloy]

@marcelofabri Keep us posted of your communication please! ❤️

[marcelofabri]

Here's what I was able to measure on our project about total images loading time:

	First Time	Second Time
Dynamic frameworks	7.3 seconds (91.0%)	66.74 milliseconds (9.5%)
Accelerator	7.7 seconds (90.6%)	43.50 milliseconds (7.8%)

[johnno1962]

Bother.. In the finish I only see the speedup on static linking using Accelerate in the debugger too. Well, at least this is one approach we can eliminate!

[alloy]

@johnno1962 I love your try it and eliminate hands-on approach, thanks for your efforts and input! ❤️

[orta]

This eventually came to a conclusion in 9.3 - see the eigen issue for full details