Is it really? Scaling today means running on more than one machine (google, facebook, twitter, etc.)
That means no shared memory. He helpfully makes this distinction on his front page ("I'm mostly interested in shared-memory system, so if you are looking for information about clusters, web-farms, distributed databases and the like, it's the wrong place")
According to Google's Jeff Dean, "to Google, multi-core computers look like separate servers with really fast interconnections" (i.e. memory).
So if you are running your applications on many machines anyway, you might as well drastically simplify your code by writing it "single-threaded" and running #cores copies on each machine.
> Scaling today means running on more than one machine
Of course it's not. Hundreds of millions of people use just a single computer for a lot of tasks.
>you might as well drastically simplify your code by writing it "single-threaded" and running #cores copies on each machine.
It's not the worst approach. However there may be significant penalties in terms of performance and latency in some contexts. You are definitely don't want to use your approach for games and browsers. As for server software it depends on performance/latency requirements. For example, if you will use it in High-Frequency Trading marker, count you loose all your money.
I would add -- if you're expecting to make money off this, you're probably better off advertising your own consulting services -- at suitably high rates. You're the expert now... :-)
So far, nobody has mentioned the author's C++ based race detection tool. Conceptually it is similar to Corensic's commercial product, Jinx (http://www.corensic.com) but practically Jinx supports more languages, doesn't require recompilation, and is most likely much faster.
It's "similar" to a lot of software out there: Intel Thread Checker, Chord, Zing, Spin, RacerX, CheckFence, Sober, Coverity Thread Analyzer, CHESS, KISS, PreFast, Prefix, FxCop.
However, what you are missing is that most of these tools (and Jinx as far as I see, can't find clear description on the site, mostly vague marketing stuff) are of help to you if you are an application developer who writes in term of mutexes. While RRD is of help to you if you are implementing mutexes itself. Can you verify involved mutex algorithm down to possible memory access reorderings? I doubt.
> and is most likely much faster.
Or an order of magnitude slower.
I do not see anything about memory fences. If Jinx does not support them, then it's pretty much useless for verification of synchronization algorithms. I've implemented dozens of advanced synchronization algorithms, and I may say that it's crucial.
Also, if it works on binary level (does not require re-compilation), then it also renders it useless, because on that level you lose information about order of memory accesses, memory fences, atomicity. For example, if you see plain x86 MOV instruction, what is it? non-atomic store? atomic relaxed store? atomic release store?
Actually I'm not sure if it was supposed to be funny or serious. I see the funny "everything-free" list, as well as can imagine that there is some action you can do not atomically (relative to other actions) that gives you synchronisation.
I'm not quite sure what you mean, but synchronization without atomic operations is possible.
An example of mutual exclusion, without any atomic operations, taken from the book "The art of multiprocessor programming"[1] is (paraphrased) as follows:
Two threads, A and B, want to access some memory. Each thread has a flag.
When thread A wants to access the shared memory:
Set flag A
Wait for flag B to become unset
Access memory
Unset flag A
When thread B wants to access the shared memory:
Set flag B
While flag A is set {
Unset flag B
Wait for flag A to become unset
Set flag B
}
Access memory
Unset flag B
Obviously this isn't a general purpose solution, but rather an easy to understand example demonstrating that atomic operations are not required.
I chose that example because its easy to understand, obviously in modern processors with out of order execution and whatnot, you would need something a lot more elaborate.
Once you add the appropriate memory barriers, it looks a lot more "atomic"
Well, they force in order memory access. That doesn't look terribly "atomic" to me, but I understand your point.
The intention was actually something like "atomic-RMW-free", that is "costly-operation-free". It should also include expensive memory fences, and excessive sharing.
I have an explanation of "atomic-free" in the Russian article, but I did not transfer it to the English page on xxx-free. Sorry for any confusion. I think I better just remove it from the main page for now.
