Meta's Transparent Memory Offloading Saves Them 20~32% Of Memory Per Linux Server

[phoronix]

Phoronix: Meta's Transparent Memory Offloading Saves Them 20~32% Of Memory Per Linux Server

Meta's engineering team today published an interesting blog post about Transparent Memory Offloading (TMO) as a new Linux kernel feature they developed that is already used in production on Facebook/Meta servers. Within Meta's data centers this TMO functionality is saving 20~32% memory per server across their millions of servers...

https://www.phoronix.com/scan.php?page=news_item&px=Meta-Transparent-TMO

[zexelon]

Now this is really cool! They are basically bringing some of the high level ideas behind their tectonic file system into memory! I wonder if this could eventually lead to a better implementation of the swap mechanism in Linux?

[JEBjames]

Michael

Typo/grammar "already an then" should probably be "already and then"

[chuckula]

This is basically a better swap algorithm, and I have no idea how much work has been going into Linux's swap algo in the last couple of years, so it could definitely be helpful in situations where the default swap isn't cutting it.

[Jabberwocky]

I don't think I'll ever get used to the name...

Not because the name changed, also because the word "meta" is such a bad word. Language is not supposed to confuse people. We use it to help us communicate. Concepts like free vs libre already confuse people. The word meta is on a whole new level. It's too confusing and overused before Facebook adopted it. People talk about things like the metaverse, since they took the new name do they now own copyright/trademarks to words like that?

[ehansin]

I don't think I'll ever get used to the name...

Not because the name changed, also because the word "meta" is such a bad word. Language is not supposed to confuse people. We use it to help us communicate. Concepts like free vs libre already confuse people. The word meta is on a whole new level. It's too confusing and overused before Facebook adopted it. People talk about things like the metaverse, since they took the new name do they now own copyright/trademarks to words like that?

Dude, that is so meta!!

https://www.merriam-webster.com/words-at-play/meta-adjective-self-referential

[ddriver]

Yey, hurray for the swap file. You can already get a massive improvement via large pages, ssds aren't particularly strong at 4k iops, they need larger buffers to shine

But also, just because they now want to be called meta, doesn't mean you have to do it. Did they even explain how exactly are they "meta", or are they simply changing it like a disposable diaper?

In my book, it is far better to call things as they are, not as they want to be called. So farcebook it will remain. And I don't think the company will ever merit an upgrade from that.

Farcebook has really been committing to improving the part of its software that runs on their end - their servers. Now how about addressing the utter garbage their servers throw at client browsers? Seriously, is there any front end framework more slower and bloated than react?

[CommunityMember]

saving 20~32% memory per server across their millions of servers

And given that most of Meta's servers have a huge amount of memory, that ends up being a huge amount of money.

At this point, many of the most "interesting" improvements in the kernel come from the HPC/Hyperscaler crowd who actually contribute with real money (staff). The idea that individuals matter (and that this year will be the year of the Linux desktop) is, at best, delusional.

[sarmad]

And given that most of Meta's servers have a huge amount of memory, that ends up being a huge amount of money.

At this point, many of the most "interesting" improvements in the kernel come from the HPC/Hyperscaler crowd who actually contribute with real money (staff). The idea that individuals matter (and that this year will be the year of the Linux desktop) is, at best, delusional.

Open source is not about individuals; it's about collaboration, whether it's coming from individuals or businesses. Also, there are countless interesting open source projects coming from individuals.

[ddriver]

Open source is not about individuals; it's about collaboration, whether it's coming from individuals or businesses. Also, there are countless interesting open source projects coming from individuals.

Let me fix that one real quick for you:

Open source is not about people, it's about big tech, whether it's stuff they need to oppress and exploit the serfs better or sharing something largely irrelevant to cultivate a positive image. It started as an alleged opposition to monopolies like m$, but quickly defined itself as the "cheapo collab alternative" business model to those who didn't felt like having the big tech boys hands rummaging through their pocket. Which is why toady foss, while super useful for hyperscale enterprise or mission critical solutions, is still useless to a regular consumer. The same m$ that foss was created to oppose now makes more money on foss than the entire end user collective combined.

And no, before you say it, android users are not linux users, it is google that is using linux for android, so it can in turn use people through those devices. So in a way, an android user is rather being used by linux by being used by googul.

That's one more reason to sleep better tonight, knowing a big tech company nobly contributed to foss something it need for its own specific use case.