Improve performance by increasing the buffer size to 20K

[arp242]

Reading a 5M file takes about 13 seconds. For reference: Python takes about 1.2s for the same file, Go 0.4s, and C++ about 0.25s. So there's a lot to gain!

One reason is that when reading the file in memory it only allocates 1000 bytes at a time; a simple measurement with this:

1k.toml     0.00s
10k.toml    0.00s
100k.toml   0.01s
1M.toml     0.52s
5M.toml    12.97s
10M.toml   74.58s

And some measurements for other values:

2K      1k.toml      0.00s
	10k.toml     0.01s
	100k.toml    0.01s
	1M.toml      0.36s
	5M.toml      9.30s
	10M.toml    53.68s
5K      1k.toml      0.00s
	10k.toml     0.00s
	100k.toml    0.00s
	1M.toml      0.27s
	5M.toml      7.02s
	10M.toml    44.81s
10K     1k.toml      0.00s
	10k.toml     0.00s
	100k.toml    0.01s
	1M.toml      0.25s
	5M.toml      6.27s
	10M.toml    43.05s
15K     1k.toml      0.00s
	10k.toml     0.00s
	100k.toml    0.01s
	1M.toml      0.23s
	5M.toml      5.96s
	10M.toml    39.19s
20K     1k.toml      0.00s
	10k.toml     0.00s
	100k.toml    0.01s
	1M.toml      0.24s
	5M.toml      5.66s
	10M.toml    39.41s
25K     1k.toml      0.00s
	10k.toml     0.00s
	100k.toml    0.00s
	1M.toml      0.21s
	5M.toml      5.77s
	10M.toml    38.10s
50K     1k.toml      0.00s
	10k.toml     0.00s
	100k.toml    0.00s
	1M.toml      0.21s
	5M.toml      5.38s
	10M.toml    33.91s

I set it to 20K as the performance benefits drop off after that, at least on my system, and it's still quite little memory, but can also use another value if you prefer – 5K or even 2K already make a difference.

(the rest of the performance is mostly in the strcmp()s in check_key() by the way, that loops over all the keys for every key it finds, which is why larger files get so drastically slower).

[arp242]

Can also start at 1K like before, and then grow to 20K after we've done a few loops (and after that even further): that way memory usage stays small for small files, but is still fast for larger files.

I don't really know what your expectations for memory usage are. 20K is nothing for any desktop system, but can be relatively a lot for embedded systems, I guess?

[cktan]

In my view, TOML is primary meant for configuration. Most TOML files probably have less than 100 lines. 1K seems very small, and I don't have objections to grow it to around 16k.

[arp242]

Yes, I agree. It's just that it would be nice to also make it work reasonably well with larger files, and seemed a simple enough change.

I changed it to 16K now.