size increase of core116

[r10s]

core116 is significantly larger than core112.8 - when building the new android 1.37, we got the following apk sizes (4 arch, all compressed):

android1.36.5 using core112.8:         56.3 mb
android1.37   using core116:           67.9 mb  ->  +11.6 mb compared to android1.36.5

android UI code or dependency itself did not grew significantly (first line shows is compiling new UI with old core) - but compiling 1.37 with --disable jsonrpc makes a difference:

android1.37, core112.8:                56.4 mb
android1.37, core116 without jsonrpc:  61.1 mb

part of the size increase is already detected by @link2xt already as "brotli". compiling with disabling unused brotli features as of #4461 , we get down 5.5 mb:

android1.37, #4461:                    62.4 mb  ->  +6.1 mb compared to android1.36.5
android1.37, #4461 without jsonrpc:    60.6 mb

idk, why the size decrease with --disable jsonrpc is less when using #4461 instead of core112.8.

i double checked that these sizes are not related to android UI code by compiling without jsonrpc with and without throwing out all jsonrpc depencies added at deltachat/deltachat-android#2479

maybe the size increase of then 6.1 mb is acceptable now, esp. as it maybe hard to get to the cause.

[adbenitez]

@r10s maybe interesting: DeltaLab release from 2 days ago has only + ~4MB increase (+ ~1MB per arch) and that is without the recent optimizations made by @link2xt

[adbenitez]

I remember I sometimes published DeltaLab with bigger size than expected due to some caches after several compilations, maybe try some cleaning before building

[r10s]

interesting, there is also some flakiness in the measurement above (see comment "idk, why the size decrease...") - any idea where this may come from?

wrt cleaning before building: yes, this is what i am also always doing 🧹

[link2xt]

Another improvement, setting codegen-units to 1: #4464

interesting, there is also some flakiness in the measurement above (see comment "idk, why the size decrease...") - any idea where this may come from?

Could be due to the default codegen-units = 16 and parallelization? May be the work is split differently depending on the order in which the cores finish their tasks.

There is a related upstream issue about codegen-units not being 1 for the releases by default: rust-lang/rust#47745

[link2xt]

I have just triggered a nightly build with #4461 merged in, size of the nightly is 67483126 bytes: https://download.delta.chat/android/nightly/2023-06-08/

Previous was 67951509 (https://download.delta.chat/android/nightly/2023-06-07/) and 67954600 (https://download.delta.chat/android/nightly/2023-06-06/).

So it looks as if #4461 alone is saving 0.4 MB.

[r10s]

building android 1.37 with #4464 we get a 4.6 mb smaller apk compared to default codegen-units. for better comparison with android 1.36.5, i also created a build of the old core with codegen-units=1:

android1.36.5, core112.8, codegen-units=16  56.3 mb
android1.36.5, core112.8, codegen-units=1   54.5 mb
android1.37, core116, codegen-units=16      62.4 mb
android1.37, core116, codegen-units=1       57.8 mb -> +3.3 mb compared to core112.8, codegen-units=1  

so, an increase of 3.3 mb for added jsonrpc and brotli, that's still something, but probably okay i'd say. from the original, too big, release of 1.37, we're down 10.1 mb :)

interesting, that default codegen-units increase android1.36.5 by 3% where it increases android1.37 by 8%. the new codegen-units=1 compilation is maybe better comparable.

nb: @link2xt i always call ./scripts/clean-core.sh before benchmarking/releasing, to have a clean start and not sth. in some caches that may break things - wondering if that script is sufficient?

[link2xt]

After merge of #4464 the size of nightly at https://download.delta.chat/android/nightly/2023-06-08/ is down to 62540285 bytes.

[r10s]

k, i think, we optimised the reasonable things and there is no actionable item left. thanks everyone!