So I'm using JSDB on the EV3 (which is an ARM5 running Debian Stretch).
I've been modelling a "new" Javascript API based on the existing Java API (since it is fairly easy to match the OOP from Java->Javascript). Java has a tendency to push everything out into single files for classes. Because of this, my new API is also in many classes. I did a test where I actually concatenated everything into a single file, and it runs blazingly fast (in people-time). Stream, even on the ARMv5, calling regular system functions (cat, ls, etc..) is very fast, but loading files (with "load") is very slow. I suspect this is a hardware issue due to the micro-SD used in the EV3. It probably has a similar internal architecture for storage like that of a Raspberry Pi, making any IO to the disk costly. I even bet, if there was a way I could make the original Python library a single monolithic file, it'd probably run very fast too (something worth looking into).
For the API, to answer your question, all the load()'s are happening at startup and are very costly getting things going.
The only downside I've had with JSDB is the fact that it is still on ES5 (with a few small gaps, like missing some of Object). I'd love to see a newer version that could run ES6, and maybe even the proposed ES7 that'll come in 2020. Eclipse hooks for debugging would also be a really nice addition, but alas, all of that takes lots of time from experienced people, and my knack is at transcoding APIs, not writing interpreters. So for now, I'll continue using ES5 on JSDB on EV3Dev for Lego and hopefully our team will continue doing well when the compete . I'm just trying to get them onto a platform that is both competitive and non block-programming (like the mindstorms environment).