Race Conditions == Random Number Generation

by dasmithii • 12 years ago

91 50 12 years ago

emhs• 12 years ago

Color me impressed. I do wonder if there's any way to reliably influence the pattern of the mutations. On the surface, with the differences in thread-scheduling patterns, I'd say it seems like a meaningful improvement on the randomness of the system PRNG.

If you have a sufficiently well-developed model of thread-scheduling patterns, then the randomness reduces in quality to that of the system's PRNG. But thread-scheduling is a serious pain, so this seems promising.

gh02t• 12 years ago

I'm pretty sure you can reduce the entropy of the output by saturating the processor or memory. If the various accesses are being spread out more because other processes are executing then you're going to have substantially less entropy. Sounds like bait for a DoS attack... hit the server really hard and all of a sudden the RNG is producing bad random numbers.

I doubt it's more efficient than something like Mersenne Twister. Generating entropy this way is pretty damn expensive in terms of processor cycles.

breathoftea• 12 years ago

generator.c:

    // For each byte in the buffer, use its value to index into another
    // byte and XOR the two.

This seems to be the only mixing that happens in the entire algorithm, after the initial rand seeding.

I'm not a cryptographer, but this smells extremely insecure to me. If the random seed is all zeroes this generator will generate all zeroes for perpetuity, no?

Even if you get a decent initial seed, I still strongly suspect this RNG strategy will be very statistically leaky, will tend to equilibria, and will generally be easily breakable in practice. An RGB visualization in the context of RNG means next to nothing.

Please children, don't do your own crypto.

mcross• 12 years ago

It also uses 100% CPU until the threads are stopped. The author did not make any claims that it was cryptographically secure. It's just a toy that he shared - an interesting way to perhaps generate something useful from something that is normally undesirable.

dasmithii7• 12 years ago

I suppose the 100% CPU is (very) unnecessary.

Maybe I'll include a variable sleeping time determined by current buffer values. That way, thread synchronization would be further disorganized, and CPU power wouldn't be wasted quite as much.

In that case, it's important that mutations take place between every byte outputted tough.

nathan7• 12 years ago

The visualisation means next to nothing for proving it's good. But the visualisation seems to show patterns, which says plenty.

dasmithii7• 12 years ago

It seems to be an improvement from the notoriously poor rand() of PHP.

(not confident in the credibility of this article) http://boallen.com/assets/images/randbitmap_computer.png

LnxPrgr3• 12 years ago

If I look at it too long, I see "patterns" in the output of OS X's /dev/random too: http://codeanarchy.com/files/yarrow.png

Random doesn't always look random. To make matters worse, we're biased towards seeing patterns in noise: http://en.wikipedia.org/wiki/Clustering_illusion

kyberias• 12 years ago

"Patterns"? Of course there are patterns. If there were no patterns, that would be suspicious. Do you claim to be able to look at a picture of a random number generator data and determine the quality based on that? You do realize that even perfectly random data may have patterns?

DanBC2• 12 years ago

"Patterns" doesn't mean much. Repeated diagonal lines would be obviously broken.

heavenlyhash2• 12 years ago

This is a remarkably poor idea.

Thread scheduling may be difficult to predict precisely. This is a radically different property than being random.

A good source of randomness is becomes no more predictable given previous outputs. Being able to predict the next output precisely would obviously break a random number generator. But being able to weight the odds of the next output also break a random number generator; it's less obvious, but most certainly enough to beat the house in any gamble.

Which thread wins race conditions is certainly not random. CPU cache lines will create biases, the implementation of your processor's pipelining of memory fences is likely predictable, and on and on.

Use a CSPRNG.

A counterargument might begin with "But I don't need cryptographically valid randomness!", to which I would respond "then why are you thinking about this at all?" Cryptographically valid randomness is not hard from either a programmer braintime perspective (there are libraries, unless you take up deep issues with modern CSPRNGs) or a CPU-time perspective (CSPRNGs can spit bytes faster than your disk can accept them). If you don't need cryptographic randomness, then there are other even faster sources of psudeorandomness; use those, not this mechanism for creating CPU waste heat.

sillysaurus3• 12 years ago

I don't think the author was presenting this as an alternative to a CSPRNG. (Or rather, I hope they weren't.) It seems like more of a fun hack than anything else. I wish they explicitly warned people not to actually use this, but it's still an interesting side project.

heavenlyhash2• 12 years ago

Right. As an amusement, this is great, and I can hardly improve upon the visual aid of monkeys frenzying over a lego structure.

And indeed even as an educational opportunity to explore how very much not-random CPU scheduling over unguarded memory is, this is also probably a gem.

But I should hate to see anyone mistake this concept for useful.

dasmithii7• 12 years ago

[author here]

You're correct. It's a horribly bad idea, and the thought I put into it was pretty minimal. Just a fun hack, really.

I'll put a note in the readme.

gamegoblin• 12 years ago

Using OS thread functionality to do some useful work is always amusing.

This reminds me of sleep sort. For those unfamiliar, sleepsort is where you have an unsorted array, then have 1 thread per element sleep for the amount of time specified by that element, then append it to a list.

dasmithii7• 12 years ago

Interesting.

Though sleep sort is obnoxious (& hilarious), I'm sure it could be of use in educational situations. It's both elegant and intuitive. Thought provoking, maybe?

AnthonyMouse2• 12 years ago

In reality sleep sort is not actually doing the sorting, it's just asking the OS to do it using a traditional sorting algorithm, because the OS will have to sort all the sleep timers to determine the order in which to wake up the threads.

mtdewcmu• 12 years ago

Hilarious, yes, elegant, no. There's no guarantee that the list will actually come out sorted (unless you dig through the implementation of sleep() down to the kernel and confirm that the order is always deterministic). It's funny precisely because it has nothing at all going for it, and yet almost kind of works.

dspillett4• 12 years ago

I'm sure it could be of use in educational situations.

It is a good example of how algorithm design can be done naively and algorithm analysis can be done badly, so definitely has at least some educational value.

Excluding implementation specific edge cases that would break the underlying assumptions regarding thread wakeup times, it is a O(N) time complexity sort in all (best, worst, average) cases. A miracle! Get thee to the patent office! Eat my shorts Shell!

Of course the constant time multiplier involved is so massive that it would not do better than an n(log n) or even an n*n sort in any practical situation, and this is before considering space requirements and thread processing overhead.

joshribakoff• 12 years ago

Those who are pointing out that the CPU & threads operate in a deterministic manner should consider that the universe may operate under the same principles. Maybe some day a computer can watch live video of someone rolling a die & predict what it will settle on before it is settled. There may be no concept of "random", only more or less predictable to humans or algorithms they create.

dasmithii7• 12 years ago

[author here]

It's difficult to have meaningful discussion (w/ regard to randomness) when involving more philosophical principles. For those who believe wholeheartedly in a deterministic universe, the concept is utter nonsense.

That's why it's important that pragmatic folk focus on practical and statistical randomness, rather than universal randomness.

By practical, I mean that the information required to predict generated numbers is simply not available. It may exist; it may not; but regardless, it's inaccessible to those in need of it.

In this case, I'm not certain if sufficient thread-scheduling information is available to derive accurate predictions, or if separate processes could be of effect without altering the speed at which RNG state mutation occurs simultaneously. Being interdependent on other thread & process CPU usage, my approach might be another case of "the act of observation changing the thing observed".

But then again, I'm know little of both cryptography and CPU scheduling. This comment, like this project, is largely speculative.

Eridrus• 12 years ago

The issue I've always had with this & dakarand are that they might not be easily predictable, but it is probably possible to influence their output into something you can more easily predict.

pjc50• 12 years ago

There's a book "The Newtonian Casino" which describes doing something similar with roulette wheels .. in the 1980s.

DanBC2• 12 years ago

Doesn't "sensitive dependance on initial conditions" place severe limitations on what we can predict?

cpeterso• 12 years ago

My silly random number generator:

strace -Tiv -ttt nice -n 19 curl -Lv --raw https://google.com/news 2>&1 | shasum -a 512 | dd bs=1 count=2 2>/dev/null

This assumes Google News serves frequently changing content. The noncanonical Google News URL generates extra entropy from an HTTP 302 Moved Temporarily to an HTTP 301 Moved Permanently to an HTTP 200 OK with the final news page. strace and nice and curl add entropy from OS syscall timings and HTTP headers.

(OS X doesn't have strace, so you can just skip it.)

bdhe• 12 years ago

A lot of the time, the randomness used must be secret (for nonces etc.) but your solution is an interesting way to implement a CRS (http://en.wikipedia.org/wiki/Common_reference_string_model) which is very useful in some cryptographic primitives. A CRS is a common string, typically random, that anyone can access, but cannot be controlled by any one party. Another typical thought-implementation of this is using stock market prices -- predict those accurately and you'd do better things with your time than break some crypto protocols.

jokoon• 12 years ago

I'm not sure undefined behaviors are so much unpredictable...

sklivvz1971• 12 years ago

I'd find it way more interesting if it started from a non-random buffer. Like this I can't tell how much of the randomness comes from the initial seed...

ctz• 12 years ago

I once found a bad implementation of this idea in a common commercial crypto library. I calculated from observation of a bunch of runs that they were achieving about 45 bits of entropy in practice, and taking six seconds of wall time and a shitload of power in the process.

baq• 12 years ago

run it through dieharder please, just for fun.

deutronium• 12 years ago

I made a silly one that uses the GPU to generate numbers.

http://www.anfractuosity.com/projects/rndgpu/

exDM69• 12 years ago

To the OP: for additional amusement and possibly some educational epiphanies, try compiling and running this on your smartphone (assuming it's a multi core ARM).

The cache coherency rules are different in ARM and Intel CPUs, you should see quite different result from unsynchronized access to same memory locations. On ARM, you need to be more careful with explicit memory barriers, especially when you attempt to write lock-free code, or you're in kernel space.

I expect that on ARM, you're going to see a lot less random because an entire cache line written to by one core will replace another cache line written to by another core, thereby showing only the randomness contribution of one core.

atoponce• 12 years ago

How does this compare to dakarand? http://dankaminsky.com/2012/08/15/dakarand/

yuhong• 12 years ago

I think the Linux /dev/random already do this to some extent.

Race Conditions == Random Number Generation

Related Stories