[Beowulf] Tiered RAM

[Chris Samuel]

On Wednesday, 26 September 2018 1:38:15 PM AEST John Hearns via Beowulf wrote:

> I had a look at the Intel pages on Optane memory. It is definitely
> being positioned as a fast file cache, ie for block oriented devices.

Interestingly there was a (non-block device) filesystem presented this year
called NOVA targeting these sorts of NVMM devices.  LWN has a nice little
article on it (which in turn links to earlier articles on it, and the original
paper).

https://lwn.net/Articles/754505/

It's still going through rapid development and (learning from the btrfs
experience) the kernel folks aren't going to let it in without a working fsck
from the sound of things.

There's also some NVDIMM documentation in the kernel tree which is pretty
heavy going, I've just tried to skim it and I think I know less now than when
I started. ;-)

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/nvdimm/nvdimm.txt

There's also the more readable documentation for the "block translation
table" which seems to be intended to provide a way to give some atomicity
to storage transactions to NVDIMMs which are not present given the nature
of the hardware:

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/nvdimm/btt.txt

There is also the persistent memory wiki:

https://nvdimm.wiki.kernel.org/

> This worked by tiering RAM memory - ie a driver in the Linux
> kernel would move little used pages to the slower but higher capacity
> device.
> I though the same thing would apply to Optane, but it seems not.

Well the simplest way to get what you describe there might be to use
the Optane as a swap partition. :-)

Red Hat have some docs about using NVDIMMs.

https://access.redhat.com/documentation/en-us/red_hat_enterprise_linux/7/html/storage_administration_guide/ch-persistent-memory-nvdimms

Not sure I helped much there! :-)

All the best,
Chris
-- 
 Chris Samuel  :  http://www.csamuel.org/  :  Melbourne, VIC