Gabriele Svelto (gabriele.svelto.delete@this.gmail.com) on January 15, 2020 2:17 am wrote:
> Windows reserves physical space for the page file and when it's system managed it rarely exceeds
> the size of the physical memory. If you want a 1234 TiB page file you need a 1234 TiB disk because
> that won't be lazily allocated. It's not like having a disproportionate page file has no cost,
> especially since the alternative - overcommitting - doesn't force you to pay that price.

The "normal" size for system managed is 1.5 times the physical memory, and it may grow, like on my working computer where it is 3 times the size of physical memory.

> You seem not to understand how Windows memory allocation works (and BTW Linux doesn't necessarily
> use a swap partition, it can use a swap file too). Windows memory can be mapped, you reserved
> a chunk of address space, committed, you requested that chunk of memory to be backed by something
> either physical memory or swap file, or in use which means you actually touched the memory you
> had previously committed. If you commit more memory than your page file + physical memory the
> memory allocation will fail even if you haven't touched a single byte of memory.

Sure, that's exactly the behavior some are advocating here, and that's what Windows API promise, if you commit 1TB of memory Windows will only return success if it can grantee you will have access to all that memory, and will fail if it can't, so apps may choose a less memory hungry algorithm, failing to commit 1TB doesn't means it can't commit 1GB, but why assume apps will commit memory that will never be touched?

If, in your app, you want another behavior you can allocate virtual address and write an handler to commit pages as soon as they are touched, allocated but not committed do not take space in the page file.

> So what about this guy?
>
> https://crash-stats.mozilla.com/report/index/a2a937b3-4b2a-4e4c-8435-0dc6b0200108
>
> He has 64GiB of RAM and a page file twice as large. He still has 60GiB of free physical memory
> but he ran out of commit space and promptly Windows started to fail memory allocations. How
> large do you suggest the page file needs to be in that case? How is going OOM when you still
> have 60GiB of free physical memory a good idea? Or a good experience for the user?

WTF was committing 140GB of memory?