Hacker News Reader

Comment by FullyFunctional | Hacker News Reader

Absolutely batch and use asynchronous completion. All modern systems (NFSv4, GPUs, NVMe, etc) uses a variation of the theme. However, one should also work on lowering the context switch cost.

Another notion that I find interesting is one-address space. You still get per process protection, but address space is global. Supporting virtual memory is extremely expensive (we are paying the price today with hardware table walkers, multi-level caching of translations, various flushing on context switches). It is much cheaper if we only have to implement permissions (there are many options here) and it can make zero-copy process communication much cheaper.

Also, if you are giving up on paging, we can move beyond the 4096 byte page which we got with the 1962 Atlas (it had the equivalent of 96 KiB total memory; if pages had kept that ratio we would be using ~ 8 GiB pages today).

View on HN

> Another notion that I find interesting is one-address space. ... Supporting virtual memory is extremely expensive

Would it be possible to make linux work with one address space? It seems like that should be a reasonably easy thing to do given how many different hardware platforms linux runs on. And it'd be interesting to know how much performance uplift you get from not flushing as much during context switches.

Or are there more complex interactions with different parts of the kernel that I'm missing?

I didn’t assume Linux binaries would work and I don’t see how to make that work, but general application that doesn’t depend on mmap or the Linux ABI can work.

> Would it be possible to make linux work with one address space?

Not without major changes to programs' ABI. On Linux with an address space per process, programs depend on having their private resources on fixed addresses.

The upcoming/vaporware The Mill CPU offers only a single address space, and emulates fixed addresses by aliasing a fixed part of each process' address space to somewhere else — in hardware.

In software, the ABI would have to pass a pointer to each callee's context when calling them. This is e.g. what you did in AmigaOS when you called a library function. But not all functions can be this way. You don't want all "function pointers" to be fat: function and context. For those to work, dereferenced functions would need to either belong to the program binary only, be "pure" (not access any global variables) or use a system service (on a fixed address..) to look up its context.

But the benefit of virtual memory is the contiguous memory (that those virtual indirections allows for) - ignoring virtual memory we risk falling down again to the "you need 600kb [of contiguous] conventional memory" issues of fragmented memory.

A MMU is complex, but allows simpler programs in the entire system overall.

I agree that abandoning virtual memory entirely brings back old problems, but there is a middle ground of a single virtual address space. This limits flushes to page table changes and makes it easier to move the MMU out of the CPU cores/caches. Page table size issues can be mitigated by larger or mixed size pages (and already is in x86_64/ARM)

Reply on news.ycombinator.com