| > "Ah, but each container will run it's own separate runtime process." Sure, but the most valuable resource that probably wastes is a PID (and however many TIDs). Processes exec'ing the same program will share a .text and .rodata sections and the .data and .bss segments are COW'ed. In terms of memory footprint, I can allow that the OS may be smart enough to share a lot of the exec'ed process if it's run multiple times. With Go and Rust and other statically compiled programs, that's going to scale to the number of instances of a service. With Node you might scale more, but then you need to start dynamically loading app code and that won't be shared. With wasm hosts, you can just ship your app code, and ask the wasm host to provide libraries to you. So you can have vastly more memory sharing. Wasm allows a lot of what you term p_i to be shifted into k through this sharing. But there's so many other reasons to have a a shared runtime rather than many processes. Context switching can be a huge cost, one that a wasm host can potentially avoid as it switches across different app workloads. Folks see similar wins from v8 isolates, which for example CloudFlare has used on their worker platform to allow them to scale up to a massive number of ultra-light worker. > Even for deploying wasm containers. Maybe there are certain technical reasons why they needed an alternate "container" runtime (wasi) to run wasm workloads with CRI orchestration, but size is not a legitimate reason. If you made a standard container image with the wasm runtime and all wasm applications simply base off that image and add the code, the wasm runtime will be shared between them, and only the code will be unique The above talks to some of the technical reasons why an alternate runtime enables such leaps and bounds versus the container world. Memory size is absolutely the core reason; that you can fit lots of tiny micro-processes on a wasm host, and have them all sharing the same set of libraries. Disk size is a win for the same reason, that containers don't need to bundle their dependencies, just ask for them. There's a 2022 post talking about containers, isolates, and wasm, talking more to all this arch:
https://notes.crmarsh.com/isolates-microvms-and-webassembly |
The question is really if you want hundreds of big and medium-sized services on a server, or tens of thousands of tiny services. This is a design question. And while my personal preference would be for the former, probably because that's what I'm used to, I'll admit there could be certain advantages to the latter.
Good job, you've convinced me this can be valid.