Performance of window.ipfs

[lidel]

Content script is injected into every webpage, which comes at a price:

• add-on/dist/contentScripts/ipfs-proxy/ is ~2.8MB

This raises some questions:

• what is the performance impact on general browsing experience?
• how can we make it better?

cc @alanshaw

---

Status

• Cut content script run time by removing API client constructor from window.Ipfs (PR: Remove API client constructor from window.Ipfs #467)
• Minimize memory allocation by lazily establishing the IPC port from the content script to the background script

[alanshaw]

I'll measure where the bottleneck is, but some ideas:

1. We include the IPFS constructor in the page, which is ~3MB of JS that needs to be read and parsed by the browser (this was the reason for the issue with the dropdown being slow to appear).
2. All pages open a messaging "port" to the background process, even if it's not used - this could be done lazily
3. We have all_frames: true which means that all frames of a site get a window.ipfs. Imagine performance issues of 1 & 2 but 10x because the site has loads of adverts that are served in an iframe for example

My hunch is that we can mitigate 3 by solving 1 and/or 2.

[alanshaw]

My findings are as follows:

On my circa 2014 MacBook Pro the content script currently take the following time to run:

• ~1,100ms (Firefox)
• ~900ms (Chrome)

If I remove window.Ipfs:

• ~500ms (Firefox)
• ~300ms (Chrome)

Which is >50% time saving.

The problem is exacerbated on sites what have many iframes. For example, http://theregister.co.uk had 12 iframes when I ran the tests.

The content scripts are run in serial, i.e. the next one does not start until the current one is finished. It means that I can actually see the adverts load in one after the other.

I believe the time is being spent parsing the IPFS library - it's nearly 3MB of JS.

My recommendation would be to remove adding the constructor window.Ipfs.

N.B. I experimented with lazily establishing the IPC port from the content script to the background script but it has no significant effect on run time.

Also, I checked with minified scripts and it also doesn't have a significant effect on run time.

[lidel]

Thank you for looking into this!

My recommendation would be to remove adding the constructor window.Ipfs

I agree. It looks like it is barely used anyway.
If needed, developer can always just include it via <script>.

Some additional data for add-on/dist/contentScripts/ipfs-proxy/ with window.Ipfs:

-rw-r--r-- 1 lidel lidel 2.8M Apr 23 11:15 content.js

And without window.Ipfs:

-rw-r--r-- 1 lidel lidel 1.1M Apr 25 12:00 content.js

That is ~1.7MB difference (60%!)

Will create PR shortly. Edit: PR is at #467

N.B. I experimented with lazily establishing the IPC port from the content script to the background script but it has no significant effect on run time.

Just to clarify: no at all, or no effect on load time? Do we know what is impact on memory usage? I would suspect that every time extension is establishing a new port some memory goes away, and if we can avoid that, that would be great.

[alanshaw]

Just to clarify: no at all, or no effect on load time?

I haven't taken any memory measurements, but we should probably do lazy IPC port anyway as there is an overhead.

[lidel]

A note from #webextensions

Will: thinking about how executeScript probably works internally (cross-process messaging), it might be more efficient to reduce your number of executeScript calls to just a module loader script that could then load all your other scripts in parallel whilst ensuring they execute in the correct order

[lidel]

Q3 summary:

• The biggest performance issue got fixed in fix: slow window.ipfs.files.add in Firefox #582
• Injection overhead will be removed entirely in Q4 when Meta: window.ipfs v2 #589 lands ← follow future work in that issue