Hello Tom,
>>> Maybe on OpenBSD pg should switch srandom to srandom_deterministic?
>>
>> Dunno. I'm fairly annoyed by their idea that they're smarter than POSIX.
Hmmm. I'm afraid that is not that hard.
>> However, for most of our uses of srandom, this behavior isn't awful;
>> it's only pgbench that has an expectation that the platform random()
>> can be made to behave deterministically. And TBH I think that's just
>> an expectation that's going to bite us.
>>
>> I'd suggest that maybe we should get rid of the use of both random()
>> and srandom() in pgbench, and go over to letting set_random_seed()
>> fill the pg_erand48 state directly. In the integer-seed case you
>> could use something equivalent to pg_srand48. (In the other cases
>> probably you could do better, certainly the strong-random case could
>> just fill all 6 bytes directly.) That would get us to a place where
>> the behavior of --random-seed=N is not only deterministic but
>> platform-independent, which seems like an improvement.
>
> That's a point. Althought I'm not found of round48, indeed having something
> platform independent for testing makes definite sense.
>
> I'll look into it.
Here is a POC which defines an internal interface for a PRNG, and use it
within pgbench, with several possible implementations which default to
rand48.
I must admit that I have a grudge against standard rand48:
- it is a known poor PRNG which was designed at a time when LCG where
basically the only low cost PRNG available. Newer designs were very
recent when the standard was set.
- it is a LCG, i.e. its low bits cycle quickly, so should not be used.
- so the 48 bit state size is relevant for generating 32 bits ints
and floats.
- however it eis used to generate more bits...
- the double function uses all 48 bits, whereas 52 need to be filled...
- and it is used to generate integers, which means that for large range
some values are inaccessible.
- 3 * 16 bits integers state looks silly on 32/64 bit architectures.
- ...
Given that postgres needs doubles (52 bits mantissa) and possibly 64 bits
integers, IMO the internal state should be 64 bits as a bare minimum,
which anyway is also the minimal bite on 64 bit architectures, which is
what is encoutered in practice.
--
Fabien.
