rmandvikar.Random2 3.1.0-alpha0

Random

Few fast thread-safe C# Random implementations.

Background

See: https://csharpindepth.com/articles/Random

TLDR, Random in C# has 2 problems that developers have to code around:

1. Random is not thread-safe. In some cases when the Random instance is accessed by multiple threads, it could yield numbers that will always be 0 and then that instance becomes useless. This is for net framework, net core, and net (net5.0).
2. Random instances new'ed up in close time intervals will result in same numbers. This is because Random is deterministic where instances with same seed will give same numbers. The default Random ctor internally uses Environment.TickCount as seed which doesn't change for 10-16ms on net framework. This is fixed on net core, and net (net5.0).

Random API

Random doesn't inherit from an interface so the implementations inherit from Random versus a static class. This makes it easier to swap in whichever implementation that suits the need. RandomFactory gives out a singleton instance of the implementations as a single Random instance per app domain is the recommendation. The ctors are not public and a specific seed cannot be used.

LockRandom simply makes pass-through calls to the base instance's methods with a lock, so using a specific seed technically is possible but not done.

ThreadStaticRandom and ThreadLocalRandom use similar approaches where a global RNG is used to seed the threads' Random instances. A specific seed is meaningless and the base instance is unused. In previous versions, the global RNG implementations were not exactly the same so as to keep them as the original authors intended, but using a cryptographic RNG for the global seed is better.

To address the shortcomings, a static member with thread-safe impl as Random.Shared was added in net6.0. It uses ThreadStatic underneath (see pr).

Usage

Local variable

// ok to capture in a local var
var random = RandomFactory.GetThreadStaticRandom();
var random = RandomFactory.GetThreadLocalRandom();
var random = RandomFactory.GetLockRandom();

Static class member

Note that it's ok to capture the ThreadStaticRandom instance too in a static class member, as every impl is a wrapper class.

// ok to capture in a static member
private static readonly Random random = RandomFactory.GetThreadStaticRandom();
private static readonly Random random = RandomFactory.GetThreadLocalRandom();
private static readonly Random random = RandomFactory.GetLockRandom();

Tests

• Showcase_Issues: Showcases different behaviors and problems with Random approaches.
• Verify_Correctness: Showcases issue 1. with Random and how the other approaches don't have it. Particularly, Verify_Correctness_Random is the test that showcases the thread-safety issue (issue 1. above) once in a while on net framework and even net core, net (net5.0). Yikes!
• Verify_Perf: Perf 10M iterations of Next() call in parallel for different approaches.
• Verify_Distribution: Sample 100 iterations to manually check distribution (for patterns, repeats, etc).

Perf

Verify_Perf for 10M Next() calls.

Recommendations

• Use LockRandom if you want a balance of speed and to guard yourself against any issues, patterns, etc.
• Use Random.Shared directly if on net6.0+. It perfs similarly to ThreadStaticRandom impl. RandomFactory.GetSharedRandom() is added to simply provide a migration path away from this nuget.
• Use ThreadStaticRandom, ThreadLocalRandom if you want speed but are ok with the possibility of issues discovered in future (patterns, etc). They are the fastest of the bunch.
• Implementations with new Random instances on every call, and using Guid.NewGuid().GetHashCode() as seed are the slowest (issue 2. on net framework only). Guid.NewGuid().GetHashCode() could come with its own problems as it depends on Guid's GetHashCode() implementation.
• Implementation with global Random instance is riskiest (issue 1. above), so definitely avoid that.