Hacker News Reader

Comment by HarryHirsch | Hacker News Reader

Investigation reveals the truth: the code produces critically different results, even for identical starting seeds and parameters

The fellow claims to have been at Google, yet seems not to have heard about non-determinism in parallelized code? Are we supposed to throw away all molecular dynamics results that were not run on a single core?!

And what's with the beef about the mathematics? Yes, it's apparently poorly determined, but that is a feature, you can extract uncertainty from it. An engineer, even if it's a software "engineer" should have heard about experimental error!

View on HN

> heard about non-determinism in parallelized code

There's a term for it. It's "race condition". It is usually considered a bad thing. Exactly because it makes the result of the code unpredictable, unless you take special care to control it. Now they question is - did the authors of the model do it? Given that their advice is "just run it single-threaded" - no they didn't.

In MD, you can eliminate race conditions like a computer scientist would or you can listen to the underlying physics and ignore them when appropiate. In real life, we depend on weak causality, not on strong causality. A throw of dice will demonstrate the fact.

A throw of dice will demonstrate the fact that we shouldn't care about bugs in models caused by race conditions? How does that even make sense?

The "race condition" is that the random numbers fed to each thread are nondeterministic given a specific seed. Even though a specific run can't be reproduced due to the nondeterminism, the random values aren't causing math bugs.

Let's say we are doing a stock market simulation. You have 2 shares of company A and 5 shares of company B. Your profit depends on the change in share price.

    srand(seed); // generates: [rand_0, rand_1, ...]
    total_profit = 2 * profit(rand()) + 5 * profit(rand());

The race condition can result in two different outcomes given the same seed:

    total_profit = 2 * profit(rand_0) + 5 * profit(rand_1);
    total_profit = 2 * profit(rand_1) + 5 * profit(rand_0);

Both of those outcomes are valid. Even though there's a reproducibility bug, it doesn't affect the correctness of the model.

If it is significantly non deterministic based on the number of cores, then yes, it should be tossed.

Reply on news.ycombinator.com