Random: add seed parameter

Add a parameter (at configuration) that modifies the random number generated.

I am simply shifting the ID (and/or timeinteger) by a constant number.

Should I add a warning when this shift is non zero?

You can compile the code with the following way: ./configure --with-random-shift-time=12 --with-random-shift-id=122

I am not a big fan. Not any number is a good number to add. I think we should come up with a better thing.

I see two other possibilities:

• Changing the 2 constant bytes, but I am not expecting a large difference as we are only changing 2 bytes over 18.
• Shuffle differently the bytes (e.g. shifting the bits in the initial buffer buff by a given amount or any other kind of permutation).

What do you think about that? I am only shifting the buffer by a given number of bits. My main concern is that we only have 31 different random seeds, but I am not expecting to run a large number of simulation with different seeds.

  /* Start by packing the state into a sequence of 16-bit seeds for rand_r. */
  uint16_t buff[9];
  id += type;
  memcpy(&buff[0], &id, 8);
  memcpy(&buff[4], &ti_current, 8);

  /* The inputs give 16 bytes of state, but we need a multiple of 6 for the
     calls to erand48(), so we add an additional aribrary constant two-byte
     value to get 18 bytes of state. */
  buff[8] = 6178;

  /* Shift the bits */
  int shift = 5;  // TODO define it in the configure.ac
  uint16_t tmp = shift == 0 ? 0 :
    *buff >> (8 * sizeof(uint16_t) - shift);

  *buff = (shift == 8 * sizeof(uint16_t)) ? 0 :
    *buff << shift;

  *buff = *buff + tmp;

  /* Shuffle the buffer values, this will be our source of entropy for
     the erand48 generator. */
  uint32_t seed16 = 0;
  for (int k = 0; k < 9; k++) {
    seed16 ^= buff[k];
    inl_rand_r(&seed16);
  }
  for (int k = 0; k < 9; k++) buff[k] ^= inl_rand_r(&seed16) & 0xffff;

  /* Do three steps of erand48() over the state generated previously. */
  uint16_t seed48[3] = {0, 0, 0};
  for (int k = 0; k < 3; k++) {
    for (int j = 0; j < 3; j++) seed48[j] ^= buff[3 * k + j];
    inl_erand48(seed48);
  }

  /* Generate one final value, this is our output. */
  return inl_erand48(seed48);

Diving in here uninvited :)

So first of all, you're not shifting, you're rotating. I think the canonical way to do this is

const int safe_shift = shift & (1 - (1 << (8 * sizeof(buff[0])));
buff[0] = (buff[0] << safe_shift) | (buff[0] >> (8 * sizeof(buff[0]) - safe_shift));

The safe_shift is just making sure we don't shift by more bits than we have (assuming this is a power of two).

Instead of rotating, though, you could also xor (^) with an arbitrary value. That's by far not as dangerous as adding a value.

Thanks for diving in.

So you mean just doing a *buff ^= RANDOM_SEED?

In any case, we need to be able to demonstrate that the generated sequence of numbers is random enough when the "seed" is changed.

There is two different tests about the random number generator in the repository. Is it enough or do you need additional tests?

@lhausammann, yes, essentially that. Since the xor just flips bits, there's no way it can reduce entropy (bold statement).

added 1 commit

• 549f4f5f - Random Seed: use xor

Compare with previous version

added 1 commit

• 6b11d9fb - RandomSeed: apply xor only on the short int

Compare with previous version

changed the description

added 1 commit

• 62f2fb24 - Format

Compare with previous version

No need to change the signature of the function. The configuration help string should say that this is a short int and ideally check somewhere that it is.

• Can you check that it does indeed pass the tests we have for a few seed values?
• Should we write the seed in the snapshots?
• Can the online documentation be amended?

Oh, and you should run both tests. Be aware that one of them takes a few hours...

added 1 commit

• de584e81 - Add doc, change random seed name

Compare with previous version

I tried a few different seeds and it seems to be correct. Pedro, are you sure about few hours? I ran the two tests and they took about 10min.

Yes it should be written in the snapshot (done in my last commit).

I have also updated the doc.

added 1 commit

• 8a9e2d1e - Add short int in the description + format

Compare with previous version

resolved all discussions