Many of those 90's GC algorithms complained of a lack of a cheap instruction for write barriers, and some were specifically about amortizing write barriers by using coarser grained ones. Azul's pivot to software only solutions came on the shoulders of Intel's VT instruction set, which allows, among other things, for nesting of virtual memory. The Azul JVM essentially behaves as a VM.

[ETA] and some of the other GCs now leverage the vast open space of 64 bit addressing to alias memory addresses to give references 2 or 3 flavors, and use that to coordinate parallel GC activities.

Zip was an ahead-of-time compression algorithm until HTTP 1.1, and so the draw was storage and transmission bandwidth, but the deciding factor was decompression time and space complexity.

Now we've been using it for online compression so long, and bandwidth is a secondary concern. Eventually someone had to take the other constraints seriously. We made LZO and friends but the results weren't that exciting, and I think a lot of people wandered off for another 10 years to work on other things, until bandwidth became a bigger concern for producers than consumers.

For containers, I'd love to talk to some mainframe OS designers about how much of containers and The Cloud are merely democratizing techniques that were proven out on big iron 30 years ago. I suspect the answer is "a lot".