AI-Powered / Machine Learning Linux Performance Tuning Is Now A Thing

[Cape]

Nice! We are one step closer to the miracle cloud-based, AI tuned, EC2 hosted blockchain mining system. Crypto is really the future of the woodworking industry!

[Yndoendo]

Years ago, people prefixed everything with "smart" like smart phone or smart TV, which is not smart at all. Now they prefix it with "AI powered", which is just another joke.

Prefixing Smart was done even before that. Watch Wolf of Wall Street and you'll see the FBI agent reading Forbes magazine with "Smart TV" on the front cover, 1991, with a picture of a CRT TV.

https://productplacementblog.com/wp-content/uploads/2015/08/Forbes-%E2%80%93-The-Wolf-of-Wall-Street-2013.jpg

[coder]

Yes, they mean something. If they actually mean something, if you know what I mean. And yes, that was intended. My point is -- In this case, there is neither a need for an actual AI-ish algorithm nor a need to actually slap that label on it. It's just an iteration through all possible configurations. If that is AI, then I've been writing AI code since I was 8.

No, AI is not simply a brute-force exhaustive search. I have no idea how many tunables there are in the linux kernel & typical device driver contingent, but it's a lot - many orders of magnitude more than you could exhaustively search. Think thread scheduling, I/O scheduling, swappiness, read-ahead, interrupt batching, I/O buffering, etc.

Dynamically adjusting these to suit your current workload is a really good idea. Kind of like VTEC for your PC!
; )

https://en.wikipedia.org/wiki/VTEC

[coder]

Forbes magazine with "Smart TV" on the front cover, 1991, with a picture of a CRT TV.

Shout out to teh V-chip. I'm sure "smart" branding of appliances and electronics goes well back before that.

Fast forward about 10 years, and I had a CRT HDTV that you could get with an embedded PC that could run a web browser. I just unplugged the tuner module (which I had instead of the PC box) and used it as a big external display for my laptop.

Anyone remember those LD players, with attached barcode readers? Ah, the 1980's, when all we needed was more lasers.

[iyxwsoekthsv]

No, AI is not simply a brute-force exhaustive search.

I never said it was. I said what they're doing here is. Or it should be. This is not the domain for AI controlled algorithms. The branching factor and therefore the number of permutations is simply not that large. It's not chess. And, look, tonnes of non-AI, human coded chess engines out there that solve millions of positions a second.

Really, server configuration is the last thing we need massive AI algorithms for. Because you can already rule out who knows how many permutations, simply because they make no sense when combined. And so forth and so on. The number of actual candidate permutations for valid configurations is really quite limited. If for no other reason then for the fact that each individual machine typically has a single workload to perform, a single task it needs to excell at. Now, if we're talking desktop use, now we're talking possible uses for more thorough analysis. Because the workload is more unpredictable.

[coder]

Really, server configuration is the last thing we need massive AI algorithms for.

"Really, warming our blood is the last thing we need to burn massive calories for", said one dinosaur to another.

[coder]

IMO, the main downside of this is how many bugs you're going to uncover by tweaking settings that people don't often change on-the-fly.

[iyxwsoekthsv]

"Really, warming our blood is the last thing we need to burn massive calories for", said one dinosaur to another.

Well, they were cold blooded and, heck, they were around for a lot longer than mammals have been so far. So, they didn't do so bad in the grand scheme of things. Yes, they went extinct but that largely is not as a result of being cold blooded. Plenty of warm blooded species have gone extinct over the course of the planet's history as well. So, calling me a dinosaur simply because I fail to see the use of AI algorithms here? Not going to fly.

Personally, I'd be more interested in auto-tuning the knobs themselves this technology is supposed to configure. As in, taking it one step further. Instead of merely changing the settings on the CPU scheduler in an automatic fashion, use that technology to develop better CPU schedulers. Now we're talking useful. And actually not being a dinosaur. Use the AI algorithms to evolve the tech, not to just configure the tech.

[coder]

calling me a dinosaur simply because I fail to see the use of AI algorithms here? Not going to fly.

You're missing the point - arguing over this is as useful as dinosaurs arguing over the superiority of mammals. In the end, any opinions they might've had about the subject were moot.

Personally, I'd be more interested in auto-tuning the knobs themselves this technology is supposed to configure. As in, taking it one step further. Instead of merely changing the settings on the CPU scheduler in an automatic fashion, use that technology to develop better CPU schedulers. Now we're talking useful. And actually not being a dinosaur. Use the AI algorithms to evolve the tech, not to just configure the tech.

Individual subsystems interact with each other in complex ways, and can't be optimally tuned in isolation. Worse yet, adapting them in isolation creates more potential for pathological scenarios to emerge. Or else they adapt too slowly and conservatively, failing to reap all the potential benefits.

If you think this is such a simple problem, you probably think HVAC control is really trivial.

https://deepmind.com/blog/deepmind-a...oling-bill-40/

[iyxwsoekthsv]

Individual subsystems interact with each other in complex ways, and can't be optimally tuned in isolation. Worse yet, adapting them in isolation creates more potential for pathological scenarios to emerge. Or else they adapt too slowly and conservatively, failing to reap all the potential benefits.

The irony in your link is astounding, really. Because it demonstrates just how much this sort of technology can achieve. Not what it cannot achieve. Furthermore, it took that same tech mere hours to leverage the fairly substantial number of variables in chess (not the least of which its fairly large branching actor) to demolish decades of manual coding, with AlphaZero needing only a few hours of self-learning to learn enough about chess to wreck Stockfish.

Yes, the technology exists to come up with superior CPU schedulers. It does not have to end at merely configuring the knobs optimally. It can be extended to actually optimizing the knobs themselves. Primarily because the arguments you made do apply to actually configuring the knobs as well; subsystem interaction and so forth. So, if the challenge is so great that optimizing the knobs is impossible, configuring them should be as well. And yet, here we are... talking about a service which employs deep learning algorithms to actually do the configuration optimization. Hmmm... I'm beginning to see a pattern, a possibility...