>
memory access was not that much slower than the CPU.Yes it was. 1995 was not 1965.
> This is why a significant portion of the CPU is the integrated memory controller and various levels of cache.
Microprocessor architectures available in 1995 had L1 and L2 caches. L2 was available for Pentiums, so it basically trickled to the consumer market.
These caches were small: misses were more likely, so optimizing for locality was super important.
I remember hearing that same talk in the 1990's: you don't want linked lists: think of the pointer chasing and cache misses, and the storage overhead for the pointers that wastes your precious little cache space (8K in an 80486).
> With regular strings, scanning or copying a string is in order memory access which will take advantage of pre-fetching of the cache.
The point of ropes is to handle very long strings. Scanning a long string, many kilobytes or even megabytes long, is not a simple memory access, though. If you have to copy the entire tail of the flat string to insert a character, but a rope does not, the rope can win in that case. See the motivating statements in the paper: "Common string operations should scale to long strings. There should be nopractical bound on the length of strings. Performance should remain acceptable for long strings."
There are various trade-off levels possible in ropes depending on the chunk sizes: a rope can have a small number of large pieces or a large number of small ones.
I suspect why ropes are perhaps not popular is that people don't care that much about the utmost performance in string processing. A performance hiccup in some string processing isn't going to make video skip, a game framerate lag, or a CNC cutter spoil a piece. A lot of text processing is batch processing, or else non-real-time interactive (like in a text editor). Moreover, the cases where ropes are useful (huge strings) are a niche within text processing. People often don't have long strings so they don't feel the sting of bad performance.