On the other hand, if you're a cloud provider that multiplexes tons of virtual cores on your physical hardware, I suspect anyone trying to do the sort of careful timing analysis required for these types of attacks would find themselves drowning in noise, as their processes get migrated arbitrarily between cores of hardware shared with tons of others.

It's odd, because this agrees with what I wrote, and the parent to your comment says they "fully concur", yet they are arguing that I'm incorrect. I did a poor job in communicating.

As an attempt to better clarify what I wrote: I agree with you that for the vast majority of people this specific attack is a non-issue. But, there are plenty of different timing attacks, and some of those may affect some people. It would follow then that some timing attacks should not be abruptly dismissed simply because it's classified as a timing attack.

However, my initial comment was replying to someone who wrote off the entire class of vulnerabilities, asserting that no timing attack of any variety has been used successfully. I find this a naive approach to vulnerability management. Instead of dismissing all attacks that are classified as timing attacks, vulnerabilities should be assessed for what they can do, the ease of doing it, and the potential impact of a successful attack.

But they just straight up don't run in constant time, so they're vulnerable to a timing attack across the network. That's clearly just a library bug? Like surely the dumbest part of a "constant time" algorithm is double checking that you ran for a constant wall clock amount of time?

It's... hard. A lot of the "constant cache behavior" and "constant time behavior" algorithms were written back in the day when the CPU speeds didn't change randomly on you, or at worst toggled between "idle" and "running hard." Think... oh, even the Core 2 days, really. They didn't switch that fast.

And then the hardware behavior changed out from under the algorithms, and nobody noticed. Now the throttling is far more rapid. So they may still be "constant instruction count," but that no longer implies constant time.

It's... complicated. :( And what's worse, even the people in charge of managing the complexity don't understand all the details anymore. When stuff like this surprises Intel, we've got problems.

While this won't fool timing attacks that are operating on the same machine as your process, the computation time becomes completely opaque to the network which is what the "remote" attacks are built on.