Efficient Seeding of Random Number Engines

Document number: P0205R1
Date: 2021-06-04
Project: Programming Language C++
Audience: Study Group 6 (Numerics), Library Evolution Working Group, Library Working Group
Reply-to: Moritz Klammler <> (OpenPGP: 2732 DA32 C8D0 EEEC A081 BE9D CF6C 5166 F393 A9C0)

Table of Contents

  1. 1 Introduction
  2. 2 Motivation and Scope
    1. 2.1 Background
    2. 2.2 The Problem with the Current Standard Library
    3. 2.3 The Proposed Solution
  3. 3 Impact on the Standard
    1. 3.1 Refinements and Additions to the Standard Library
    2. 3.2 Clarification on Random Number Engine Requirements
    3. 3.3 Deprecation of Existing Features
  4. 4 Design Decisions
  5. 5 Alternative Solutions
  6. 6 Implementations
  7. 7 Proposed Wording
  8. Acknowledgments
  9. Changelog
  10. References

1 Introduction

The purpose of this proposal is to make properly seeding a random number engine in a (non-)deterministic way easy and fast.

In order to achieve this, the following changes are proposed:

The changes are non-breaking and can be implemented as a pure library solution with current C++20 features.

2 Motivation and Scope

2.1 Background

C++11 introduced a powerful random number generation library [rand] under the <random> header. It provides random number engines [rand.req.eng] that can be used either directly as a source of uniformly distributed pseudo random integers of unsigned type or together with random number distributions [rand.req.dist] to produce pseudo random numbers according to a variety of distributions.

If an engine has an internal state of N bits, it can produce at most 2N different sequences and the longest sequence it can produce will have a period of at most 2N values until it necessarily starts repeating itself. Applications that wish to produce high-quality pseudo random numbers will therefore choose an engine such as the Mersenne twister (e.g. std::mt19937) with a large internal state. Choosing an engine with a large internal state helps ensuring that the period of the generated sequence will be large. However, in order to ensure that the number of different sequences the engine can produce is also exponential in the size of its state, it has to be made sure that the initial state is evenly distributed across all possible states. This is done by seeding the random number engine. If each of the 2N states should be chosen with equal probability as the initial state, seeding requires 2N bits of entropy.

The standard library provides the type std::random_device [rand.device], a uniform random bit generator [rand.req.urng] that is supposed (but, unfortunately, not required) to produce a non-deterministic sequence of uniformly distributed integers of type unsigned int. The natural choice for an application that wishes to generate pseudo random numbers and does not need (and often doesn't want) reproducibility is to use it for seeding a random engine that is then used to produce pseudo random numbers. In addition, the standard library provides the type std::seed_seq [rand.util.seedseq] which implements the seed sequence [rand.req.seedseq] requirements (and is the only type in the current standard library doing so). It can be used to seed a random number engine using entropy obtained from an arbitrary sequence of integers (e.g. a user-provided byte string read from a file or an environment variable to allow reproducing experiments with the same seed).

2.2 The Problem with the Current Standard Library

Unfortunately, the current standard library does not provide any convenient way to use a std::random_device to properly (in the sense that each initial state is equally likely) seed a random engine. Nor does it make doing so with a custom uniform random bit generator or writing a custom seed sequence easy.

The naïve approach that most people seem to use is drawing one random number from std::random_device (or whatever source of entropy they have) and then use the constructor which takes a single integer as seed.

template <typename EngineT>
void seed_non_deterministically_1st(EngineT& engine)
{
    std::random_device device{};
    engine.seed(device());
}

This code is severely flawed. If EngineT is std::mt19937, it has a state size of 19 968 bits. However, if an unsigned int is 32 bits (as is the common case on many platforms today), then of the up to 219 968 states, at most 232 (that is one 2−19 936-th) can possibly be chosen!

To illustrate that this is in fact a real problem, O'Neill [1] has pointed out that a std::mt19937 engine seeded like above can never produce certain integers as its first value. This can have bad consequences on real-world programs. A program that incorrectly assumes that an impossible number will eventually be produced as the engine's first (or n-th) output is broken in a very subtle way. After all, it would take extensive and practically unrealistic unit testing to do an exhaustive search over the possible seeds and even detect the bug.

In addition to seeding an engine with an integer, the standard library also provides a way to seed it with a so-called seed sequence [rand.req.seedseq]. A seed sequence may be constructed in a deterministic way from zero or more integers that provide some initial entropy. The numbers are then possibly scrambled and stored in some internal state of the seed sequence which can be externalized using the param member function. Such a memento may then be used to re-create a seed sequence of the same type in the same state. A seed sequence provides a member function generate that takes a pair of random access iterators and assigns a uniformly distributed unsigned 32 bit integer to each element in the range denoted by the iterator pair. The standard library provides a single implementation of a seed sequence in std::seed_seq [rand.util.seedseq]. This type can be used to seed a random engine more thoroughly.

template <typename EngineT, std::size_t StateSize = EngineT::state_size>
void seed_non_deterministically_2nd(EngineT& engine)
{
    using engine_type = typename EngineT::result_type;
    using device_type = std::random_device::result_type;
    using seedseq_type = std::seed_seq::result_type;
    constexpr auto bytes_needed = StateSize * sizeof(engine_type);
    constexpr auto numbers_needed = (sizeof(device_type) < sizeof(seedseq_type))
        ? (bytes_needed / sizeof(device_type))
        : (bytes_needed / sizeof(seedseq_type));
    std::array<device_type, numbers_needed> numbers{};
    std::random_device device{};
    std::generate(std::begin(numbers), std::end(numbers), std::ref(device));
    std::seed_seq seedseq(std::cbegin(numbers), std::cend(numbers));
    engine.seed(seedseq);
}

This code has a number of problems.

The preceding discussion used std::random_device as an example as it is the go-to solution for obtaining system entropy for a non-deterministic seed on C++ today. However, since nothing special about std::random_device was made use of, the same discusison applies to any uniform random bit generator.

Unfortunately, writing one's own seed sequence isn't a great solution either. First of all, doing something like this shouldn't be necessary for a common use case. Second and more severely, the seed sequence requirements [rand.req.seedseq] mandate a constructor which takes a pair of iterators, another constructor with takes an initializer list and finally a size and a param member function that are supposed to externalize the object's state such that it can be reconstructed again later. While this makes sense for std::seed_seq which is deterministic and always constructed from a sequence of integers, it makes no sense for a seed sequence that obtains its values from a system's entropy source. The best way to deal with these functions is to either not implement them (which isn't strictly valid C++20) and hope that the standard library won't enforce the type requirements strictly or implement them to unconditionally throw an exception, which is a sad and annoying option.

2.3 The Proposed Solution

With this proposal adopted, properly seeding a random engine could be as simple as this.

template <typename EngineT>
void seed_non_deterministically_3rd(EngineT& engine)
{
    std::random_device device{};
    std::seed_adapter adapter{device};
    engine.seed(adapter);
}

There would be no unnecessary copies, no unneeded tempering, no dynamic memory allocation, no introduced bias and little chance to get anything wrong on the user's side.

The std::seed_adapter would be a class template (the example above using CTAD) that wraps a reference to any uniform random bit generator and provides a generate member function which calls through to the generator's call operator. If P1068R4 were to be adopted and EngineT implements the proposed uniform vector random bit generator concept, the implementation of std::seed_adapter<EngineT>::generate would make exactly one call to EngineT::operator(). Even if that proposal should not be adopted, a standard library implementation would be free (and encouraged) to optimize this case for its own implementation of std::random_device at least.

3 Impact on the Standard

This proposal could be implemented with very little changes, none of which would be breaking anything. No core language changes are needed.

3.1 Refinements and Additions to the Standard Library

The requirements on the type parameter for a random engine's seed member function (and the corresponding constructor) should be relaxed such that it only requires the generate member function from seed sequence. Bolas [2] has argued that even today, a conforming implementation isn't allowed to call anything but generate and perhaps size due to existing complexity requirements. Unfortunately, it may still enforce their presence via type checks.

To express this relaxed requirement, a new named requirement seed generator and a corresponding std::seed_generator concept are proposed. Only a generate member function would be required. The existing seed sequence named requirement would be defined as a refinement of it and a std::seed_sequence concept added for completeness.

A new class template std::seed_adapter is proposed that implements the std::seed_generator concept and acts as an adapter for making any uniform (vector) random bit generator usable as the source for seeding any random number generator. The adapter would be non-owning (note that std::random_device is not copyable and copying a (deterministic) random number engine would generally be undesirable because the state transitions would be lost, introducing bad randomness in a way that is not necessarily obvious to users.)

This proposal has no dependencies on any other proposal. However, if P1068R4 were adopted, there would be optimization potential for the implementation of the proposed std::seed_adapter. The currently proposed wording was chosen to not preclude this optimization but it cannot mandate it either.

The author is not aware of any other proposal that depends on or conflicts with this proposal. In particular, it is orthogonal to N3547. (However, the sample implementation of std::randomize in that paper could benefit from the feature suggested in this proposal.)

3.2 Clarification on Random Number Engine Requirements

Currently, table 94 [tab:rand.req.eng] defines the various overloads of the seed member function of a type E that meets the requirements of random number engine in terms of the corresponding constructor and equality operator. For example, e.seed(q) is defined to have the effect that post: e == E(q) and complexity same as E(q). This is unfortunate because for a seed sequence q, two successive calls to q.generate (even for ranges of identical size) do not have to produce the same sequence of numbers, even though std::seed_seq happens to behave that way. The requirements for seed sequence don't mandate this; the respective row in table 93 [tab:rand.req.seedseq] says about q.generate(rb, re) (emphasis added):

Does nothing if rb == re. Otherwise, fills the supplied sequence [rb, re) with 32-bit quantities that depend on the sequence supplied to the constructor and possibly also depend on the history of generate's previous invocations.

Therefore, the author believes that it is not intended by the current standard that the following code should be guaranteed to work.

template <typename RandEngT, typename SeedSeqT>
void test(RandEngT& engine, SeedSeqT& seedseq)
{
    engine.seed(seedseq);
    assert(engine == RandEngT{seedseq});  // might fire
}

Regardless of whether this proposal will be adopted, the wording in table 94 should be updated to make it clear that calling seed puts the engine into the same state as calling the corresponding constructor but not create the impression that the assertion in the above example would necessarily hold.

3.3 Deprecation of Existing Features

It is not proposed that any existing library features be deprecated.

Possible candidates for deprecation could be the overload of E::seed that takes a single value of type E::result_type and the corresponding constructor of a random number engine E. Deprecating these overloads might be sensible because those functions actually encourage poorly seeding random number engines. On the other hand, they remain useful in situations where people want to hard-code a specific seed (such as in std::minstd_rand{42}) and don't really care about uniformly choosing an initial state from the entire state-space of the engine or know that the space is sufficiently small.

4 Design Decisions

The (updated) proposal takes care to be non-breaking with existing and compatible with potential future versions of the standard library (which might include P1068R4). For the sake of backwards compatibility, the specification that the generate member function of a seed sequence and therefore a seed generator deal in 32 bit integers could not be changed.

5 Alternative Solutions

If P1068R4 were adopted, this proposal could eventually be made obsolete by instead simply adding an additional constructor and seed member function to the random number engine requirements which takes a uniform vector random bit generator as lvalue argument and initializes its internal state by calling its range operator() instead of the param member function as the seed sequence overload does.

If p is an lvalue of a type which implements std::uniform_vector_random_bit_generator and E is a type which meets the requirements of a random number engine, then the expression E(p) shall initialize the engine's state via a single call to p(f, l) with f and l being contiguous iterators with a value type that is an unsigned integer. For an lvalue e of type E, the expression e.seed(p) shall have the same effect as e = E(p).

This would be a breaking change as it would strengthen the requirements on an existing named requirement. Of course, the standard library types could implement this nonetheless and the new requirement be made optional by introducing yet another concept or named requirement.

However, the author is concerned that this solution could be confusing because of the ambiguity with the copy constructor.

Making std::seed_adapter own the underlying uniform random bit generator instead of holding a reference to it was considered but not perused for the reasons discussed earlier, namely concerns about generators that are not copyable or non-obvious undesired effects of creating such copies in cases where this is possible.

The author notes that – would this proposal be adopted – the seed sequence requirement isn't actually used by the standard library itself. Therefore, the standard might as well just not define it anymore.

6 Implementations

The implementation of this proposal would be very simple. A proof-of-concept is available online.

7 Proposed Wording

All proposed changes to the standard mentioned in this section are relative to N4861.

Add to the synopsis of the <random> header [rand.req.synopsis] the following three declarations.

template <class S>
concept seed_generator = see below;

template <class S>
concept seed_sequence = see below;

template <class U>
class seed_adapter<U>;

Add a new section before the existing § 26.6.2.2 [rand.req.seedseq] to define seed generator.

Seed generator requirements [rand.req.seedgen]

A seed generator is an object that produces a requested number of unsigned integer values i with 0 ≤ i < 232. The generated numbers may be obtained from a non-deterministic source of entropy.

A class S satisfies the requirements of a seed generator if the expressions shown in the below table are valid and have the indicated semantics and if S also satisfies all other requirements of this section. In that table and throughout this section:

Expression

Return type

Pre/post-condition

Complexity

S::result_type

T

T is an unsigned integer type of at least 32 bits.

compile-time

q.generate(rb, re)

void

Does nothing if rb == re. Otherwise, fills the supplied sequence [rbre) with 32-bit quantities in a possibly non-deterministic way.

Ο(re - rb)

In the existing section § 26.6.2.2 [rand.req.seedseq], change the beginning of the second paragraph as follows.

A class S meets the requirements of a seed sequence if it satisfies the requirements of a seed generator and in addition, the expressions […]

In the current table 93, remove the first (S::result_type) and the fifth (q.generate(rb, re)) row which will already be covered by the seed generator requirements.

Replace section 26.6.2.4 [rand.req.eng] paragraph 4.4 by the following sentence with the appropriate cross-reference to the section defining the requirements of seed generator.

q is an lvalue satisfying the requirements of a seed sequence generator;

Add the following two new concepts for std::seed_generator and std::seed_sequence at appropriate locations.

template <
    class SeedGenerator,
    class RandomAccessIterator = add_pointer_t<typename SeedGenerator::result_type>
>
concept seed_generator = __unsigned_integral_least32<typename SeedGenerator::result_type>
    && random_access_iterator<RandomAccessIterator>
    && __unsigned_integral_least32<typename iterator_traits<RandomAccessIterator>::value_type>
    && requires (SeedGenerator& s, RandomAccessIterator f, RandomAccessIterator l) {
        { s.generate(f, l) } -> same_as<void>;
    };

template <
    class SeedSequence,
    class RandomAccessIterator = add_pointer_t<typename SeedSequence::result_type>,
    class InputIterator = RandomAccessIterator,
    class OutputIterator = RandomAccessIterator
>
concept seed_sequence = seed_generator<SeedSequence, RandomAccessIterator>
    && default_initializable<SeedSequence>
    && constructible_from<SeedSequence, InputIterator, InputIterator>
    && constructible_from<SeedSequence, initializer_list<typename SeedSequence::result_type>>
    && input_iterator<InputIterator>
    && __unsigned_integral_least32<typename iterator_traits<InputIterator>::value_type>
    && output_iterator<OutputIterator, typename SeedSequence::result_type>
    && requires (const SeedSequence& s, OutputIterator out) {
        { s.size() } -> same_as<size_t>;
        { s.param(out) } -> same_as<void>;
    };

For better readability, the above code uses the hypothetical concept __unsigned_integral_least32 which could be defined like this.

template <typename T>
concept __unsigned_integral_least32 = unsigned_integral<T> && numeric_limits<T>::digits >= 32;

Add the following new section.

Class template seed_adapter [rand.req.seedadapt]

template <uniform_random_bit_generator U>
class seed_adapter
{
public:
  // types
  using result_type = typename U::result_type;

  // constructors
  explicit constexpr seed_adapter(U& gen) noexcept;

  // generating functions
  template <random_access_iterator It>
    void constexpr generate(const It f, const It l)
    requires __unsigned_integral_least32<typename iterator_traits<It>::value_type>;

private:
  U* m_gen;  // exposition only
};
explicit constexpr seed_adapter(U& gen) noexcept;

Effects: Stores the address of gen in m_gen.

template <random_access_iterator It>
  void constexpr generate(const It begin, const It end)
  requires __unsigned_integral_least32<typename iterator_traits<It>::value_type>;

Preconditions: The address stored in m_gen (still) refers to a valid object.

Effects: Fills the range [begin, end) with 32 bit values obtained from *m_gen.

Complexity: Ο(endbegin)

Restructure the first seven rows of the current table 94 as follows. It is believed that this preserves the intended meaning of the current wording but avoids confusion.

Expression

Return type

Pre/post-condition

Complexity

E()

Creates an engine with the same initial state as all other default-constructed engines of type E.

Ο(size of state)

E(x)

Creates an engine that compares equal to x.

Ο(size of state)

E(s)

Creates an engine with initial state determined by s.
Note: For engines with an internal state larger than sizeof(s), there will be impossible states that cannot be created using this constructor. — end note ]

Ο(size of state)

E(q)

Creates an engine with an initial state that depends on a sequence produced by one call to q.generate.
Note: Depending on the behavior of q.generate, this operation might not be deterministic. — end note ]

same as complexity of q.generate called on a range of the size of the engine's state

e.seed()

void

Postconditions: e == E(). Equivalent to e = E().

same as E() not worse than e = E()

e.seed(s)

void

Postconditions: e == E(s). Equivalent to e = E(s).

same as E(s) not worse than e = E(s)

e.seed(q)

void

Postconditions: e == E(q). Equivalent to e = E(q).

same as E(q) not worse than e = E(q)

Acknowledgments

Melissa O'Neill has written a series of remarkable blog posts that discuss the problems with seeding random engines in-depth. Although not the initial motivation for the author of this proposal, that blog post provided valuable theoretical and experimental support and the fact that its author had independently suggested basically the same (revision 0) addition to the standard library was very affirming.

The discussions with the principal author of <random>, Walter Brown, were very enlightening and helped the author a lot figuring out the final details of the proposal (revision 0).

Nicol Bolas, Zhihao Yuan and Seth Cantrell have provided valuable feedback on the std-proposals@isocpp.org mailing list (revision 0).

Baum mit Augen's shared experience of struggling to properly seed a std::mt19937 from a std::random_device and the following discussion with Deduplicator (out of which came an earlier version of the code for seed_non_deterministically_2nd) were part of the motivation for writing this proposal [6].

Jonathan Wakely has provided valuable feedback regarding the use-case of custom seed sequence implementations and pointed the author to the current state of this proposal in the library evolution working group.

Changelog

The following changes were made compared to revision 0 of this proposal:

The feedback given by LEWG regarding the relaxation of the restriction to 32 bit integer types was not applied (yet) because the author does not know how this could be done in a backwards-compatible (with C++11 to C++20) way. We would have to strengthen an existing named requirement for that.

References

  1. Melissa O'Neill, C++ Seeding Surprises. 2015-04-16, http://www.pcg-random.org/posts/cpp-seeding-surprises.html
  2. Nicol Bolas, via std-proposals@isocpp.org, 2016-01-03, https://groups.google.com/a/isocpp.org/d/msg/std-proposals/sF-P4VE2Z3Q/u24T-g-hEgAJ
  3. Seed std::mt19937 from std::random_device in: Code Review, 2015-10-30, http://codereview.stackexchange.com/q/109260