Heterogeneous Memory Management Is Still Baking For Linux

[phoronix]

Phoronix: Heterogeneous Memory Management Is Still Baking For Linux

Jerome Glisse continues hacking on the very lengthy feature work item of adding Heterogeneous Memory Management (HMM) to the Linux kernel...

http://www.phoronix.com/scan.php?page=news_item&px=HMM-Linux-Memory-V9

[AndyChow]

It's a pretty good technology, but won't get used unless someone has the vision to push it. This is why a person like Steve Jobs is important. You need a large segment of industry to use a feature, otherwise it's a nice set of transistors that don't get any use. The promises ofISA-agnostic HSA is still non-existant. Show me code that profits from this architecture without having been specifically designed to do so. Right, it doesn't exist.

And probably, such software shouldn't exist. Software should be cross-compilable, and if target architecture has HMM, then it should take advantage of that.

Someone said "People who are really serious about software should make their own hardware". I disagree. Software should be hardware agnostic.

[bridgman]

I don't expect to see "applications written for HSA", and the same goes for HMM. The idea is that toolchains (eg runtimes & dev tools for standard languages like C++AMP) and drivers are modified to make use of the standard, then any application using a modified stack automatically benefits.

AFAIK we are doing what you want; it's just harder and takes longer than exposing a standard and saying "applications have to be written for it".

[spirit]

I hope I"ll improve memory hot-unplug (which is supported by qemu 2.4 for example). Currently it's not possible to unplug memory dimm device, if kernel memory is used on theses devices

[Brane215]

This still doesn't explain what is HMM good for. If you have HSA, for example, why would you need this ?
Or, for example, where exactly would you need this, but at the same time not have an access to that memory within adress space ?

[bridgman]

The "software visible face" of HSA is relatively high level. HMM is lower level code (mostly kernel) that could be used in an HSA implementation. If you aren't running with an IOMMUv2 (or even if you are in some cases) you basically need to be able to at least partially sync GPU page tables with CPU page tables, and HMM is an example of core plumbing for that.