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

[phoronix]

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

A year and a half ago I wrote about a start-up working on dynamically-tuned, self-optimizing Linux servers. That company is now known as Concertio and they just launched their "AI powered" toolkit for IT administrators and performance engineers to optimize their server performance...

http://www.phoronix.com/scan.php?page=news_item&px=Concertio-Optimizer-Studio

[boxie]

That would be very cool to benchmark!

[smitty3268]

Seems like a bit of a stretch to call it "AI powered". I guess we don't know exactly what it's doing behind the scenes, but it sounds like a pretty straightforward algorithm that just tests out a bunch of tunables based off some detected conditions. Not that that's a bad thing.

[mlau]

They should add "blockchain" somewhere in there too, and immediately do an IPO. As Xerox showed, that's all it takes to raise share prices these days.

[q2dg]

If it isn't open-source, it lacks all interest

[iyxwsoekthsv]

Not that that's a bad thing.

No, it's not. But, the first thing I was reminded of: https://en.wikipedia.org/wiki/QEMM#Optimize

It's not exactly a new idea in general.

[polarathene]

Seems like a bit of a stretch to call it "AI powered". I guess we don't know exactly what it's doing behind the scenes, but it sounds like a pretty straightforward algorithm that just tests out a bunch of tunables based off some detected conditions. Not that that's a bad thing.

You run whatever your workload is, perhaps in the first stage it's heavy on CPU usage as well as RAM, stage 2 might be disk I/O intensive and use GPU for quite a bit of compute, stage 3 might be a bit of a mix with high single threaded CPU usage instead of multi-threaded from stage 1, RAM might be high still but less actively changing, doing some CPU work and then offloading to disk gigabytes, CPU usage being bit low while writes and reads happen until it has enough data to continue.

I've had a similar workload to that with a prior job and several software in the pipeline(some of which have those various stages during their processing). The "AI powered" bit, is getting a large enough dataset of hardware usage/performance and other information to analyze for patterns, score them and create the labels/classifications that new input data can be matched to. So when the AI model is running, it looks at all the input data given about what's going on in the system, what is currently enabled/disabled, the workload is classified to one of the existing types in the model and it'll start tweaking the system according to what worked best for that type of workload, verify that improvements are as expected then inform the user about it.

AI that is buzzword these days is usually about that kind. They're blackboxes where the human dev part is to tell the computer what data it has to learn from, how to score what is valuable and what is not, what sort of outcome you're after etc. After a bunch of iterations of that process and training the AI, you get the blackbox of logic not written by a developer that supposedly does what you want. Probably not fun to debug, I don't have personal experience, and the only "bugs" I know of is when the human dev was at fault with how they trained the AI to learn to accomplish it's task rather than what the AI itself produced.

There is some really clever and well done use of this type of AI in projects. But also plenty of startups/companies jumping in on it for the buzzword/trendyness and other things to get funding from dumb investors. I saw one in an article recently claiming to be superior, but their comparisons were poorly researched and the actual results were very very wrong if you knew anything about the task it was delegating to AI, but to an investor with only a basic understanding of the task they might have been fooled.

This particular one, the code isn't likely to be too impressive like you mention, the effort is in the blackbox AI model that the code will utilize. I guess you could say it's a fancier/advanced type of markov chain?

[polarathene]

How are they going to go about patenting that? Is that going to affect anyone who wants to do an open-source implementation? I don't quite get what they're able to patent here, what would the patent do that prevents someone else using say TensorFlow to build their own AI model to classify workloads and apply/advise optimizations?

[polarathene]

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.

AI isn't a joke? There is some really cool stuff that it can do that is better suited to creating AI models through something like TensorFlow vs writing your own algorithms by hand(which you still kind of do anyway to get the AI model, but it's quite a different approach).

Yeah it's a buzzword that's sadly being pushed to make someone seem relevant and seek investment. Just like IoT/smart was. They do still mean something, but it's more of a categorization rather than unique selling point. Just because it's smart or AI powered doesn't mean it's good/better, it seems to get treated/used that way which is incorrect.

[iyxwsoekthsv]

They do still mean something

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.