Meta: Transparent memory offloading
source link: https://lwn.net/Articles/898454/
Go to the source link to view the article. You can view the picture content, updated content and better typesetting reading experience. If the link is broken, please click the button below to view the snapshot at that time.
Meta: Transparent memory offloading
Transparent Memory Offloading (TMO) is Meta’s solution for heterogeneous data center environments. It introduces a new Linux kernel mechanism that measures the lost work due to resource shortage across CPU, memory, and I/O in real time. Guided by this information and without any prior application knowledge, TMO automatically adjusts the amount of memory to offload to a heterogeneous device, such as compressed memory or an SSD.
The article doesn't say where to find the relevant code, not all of which is in the mainline kernel (and some of which runs in user space).
(Log in to post comments)
Meta: Transparent memory offloading
Posted Jun 20, 2022 22:47 UTC (Mon) by ssmith32 (subscriber, #72404) [Link]
If this means what I think it means, I wonder why they didn't start with this approach?
Up until I got to the end I kept wondering why they didn't just adjust the swap algorithm instead of building another layer that seems to just be second-guessing the original swap algorithm.. and/or tweak the file cache to not cache aggressively on fast storage..
Meta: Transparent memory offloading
Posted Jun 21, 2022 0:36 UTC (Tue) by MattBBaker (subscriber, #28651) [Link]
Historically the kernel has been able to ignore local apps since the kernel had more complete information about the state of the system, but once the app runtime has better information about the state of a cluster, the local kernel needs to made to mind the larger picture.
Meta: Transparent memory offloading
Posted Jun 21, 2022 4:02 UTC (Tue) by wahern (subscriber, #37304) [Link]
I guess another way to put it, perhaps some other OSs are trading away performance for reliability, and the strategy of trading away reliability for performance is finally catching up with Linux.
FaceBook (Meta)
Posted Jun 21, 2022 19:00 UTC (Tue) by Felix_the_Mac (guest, #32242) [Link]
I don't believe it is necessary to my position but I will point out that the linked article is on fb.com, is headed by a picture of an Facebook data centre and has a cookie policy in the name of facebook.
FaceBook (Meta)
Posted Jun 22, 2022 2:49 UTC (Wed) by flussence (subscriber, #85566) [Link]
FaceBook (Meta)
Posted Jun 22, 2022 3:21 UTC (Wed) by jake (editor, #205) [Link]
because Alphabet doesn't talk about Alphabet's software ... Google blogs still call themselves that ... should that change at some point, so will we ...
we like to use the names that people or organizations choose for themselves ...
FaceBook (Meta)
Posted Jun 22, 2022 13:26 UTC (Wed) by scientes (subscriber, #83068) [Link]
How meta of you.
And initially meta just just meant after until it's the famous title "metaphysics" in Aristotle gave it a special meaning.
Perhaps you should still clarify that Meta Inc. runs Facebook, because their overly generic branding of "Messenger" (which they then claim trademark over) is really an afront to the English language.
FaceBook (Meta)
Posted Jun 22, 2022 15:31 UTC (Wed) by ballombe (subscriber, #9523) [Link]
At least replace 'meta' by 'meta (the social media company formally known as facebook)' so there is no confusion to the preexisting usages of meta.
We all know the renaming goal is to sow confusion and distract of facebook trouble. LWN role is to reduce confusion, not amplify it.
FaceBook (Meta)
Posted Jun 22, 2022 18:07 UTC (Wed) by rgmoore (✭ supporter ✭, #75) [Link]
Sorry, but in this instance, they do not: the site name is engineering.fb.com, etc.
The website is indeed engineering.fb.com, but the blog calls itself "Engineering at Meta", not "Engineering at Facebook". The post talks about technologies being used at Meta, not technologies used at Facebook. This makes sense, because they are probably using the same technology across their platforms. This is part of the reason they renamed the parent: they're a single big social media company with a bunch of different brands that are making use of a lot of the same technology under the surface. When talking about that technology, it makes more sense to talk about the corporation (Meta) rather than the product (Facebook) because it's not just one product using it.
FaceBook (Meta)
Posted Jun 22, 2022 20:55 UTC (Wed) by klindsay (subscriber, #7459) [Link]
If an entity (person, organization, corporation, city, country, etc.) changes their name, it makes sense to me to refer to the entity by the name that they've chosen for themselves. To not do so, based on one's perception of the entity's intent, and how that perceived intent aligns with you own values, seems to me like a recipe for inconsistency and disrespect.
FaceBook (Meta)
Posted Jun 22, 2022 22:02 UTC (Wed) by mpr22 (subscriber, #60784) [Link]
(And yes, I do distinguish between that class of entity, and the other collective entities on your list.)
Meta: Transparent memory offloading
Posted Jun 22, 2022 13:45 UTC (Wed) by dschatzberg (subscriber, #130997) [Link]
> The article doesn't say where to find the relevant code, not all of which is in the mainline kernel (and some of which runs in user space).
The vast majority of the code (PSI, reclaim improvements) has been upstreamed in the Linux kernel. The only piece which hasn't been (and isn't critical) is the memory.reclaim interface (one can use memory.high instead). We have ongoing discussions about upstreaming memory.reclaim as well: https://lwn.net/Articles/892328/
Our userspace logic (Senpai) is a bit tied to our internal infrastructure, but we've published a reference implementation here: https://github.com/facebookincubator/senpai - it works on upstream kernels.
Recommend
About Joyk
Aggregate valuable and interesting links.
Joyk means Joy of geeK