Re: [Texmacs-dev] Darcy's plan for GNU TeXmacs 2.1.1

[Joy Yang Qiping]

Hi Joris,

> About plans for Scheme.  After the benchmarks, I got somewhat confused.
>
> If S7 were systematically faster than Guile (at least than Guile 1.8),
> then I think that moving to S7 is the simplest option for us.
> Otherwise, we should keep playing with Guile 3 too.
>
> If my understanding is correct, then the main problem is the LaTeX export.
> Does anyone understand why this is so much slower with S7?
>
> Another decision we could make is that we wish to keep the support
> for various versions of Scheme.  But then we would need a branch that
> works for both S7 and Guile 3.


Moving to S7 may not be as simple as we would like. Here is another issue that 
troubles me. Over my own usage experience, TeXmacs under S7 could consume more 
memory than I would like (over 1GB for single document editing).

I think S7 will benefit from a more sophisticated gc implementation. From what 
I have seen through the code, they are using the same memory pool for every 
scheme object, which isn’t ideal for TeXmacs. Since TeXmacs not only has a 
linear growth of persistent object, e.g., edit history for instance, during 
editing, but also has a lot of side computations that claims some memory for a 
short duration. This means the gc will often get triggered and are slower and 
TeXmacs will benefit from having several levels of memory pools (from 
persistent to short-lived) to adapt to different memory usage pattern. The 
memory pool grows exponentially. I have tried to tune its exponential parameter 
from 2 to 1.2~1.5. But from what I have observed, it only delays the 
inevitable. The current S7’s gc implementation cannot guarantee long-term 
stability of TeXmacs — memory consumption will grow to a very high level. I 
don’t think it is a good idea to go through with S7 as is. I could try to turn 
that into a linear growth, which should work fine for TeXmacs in most cases 
(I’m guessing). But it really isn’t meant to be a general purpose scheme 
implementation with those changes. We could also find ways to reduce those 
persistent memory throughout editing. I might also try to look for some 
parameters within TeXmacs that affect those (how many editing history to keep 
for instance) and see if it helps.

I think it would be better to keep both S7 and Guile 3 working for now. Right 
now the main problem with Guile 3 is Windows support. Neither is the perfect 
solution for us :(

Best regards,
Joy