CryptoRandom

by Stan Drapkin

CryptoRandom : Random

• .NET Random done right
• [Fast], [Thread-safe], [Cryptographically strong]: choose 3
• Subclasses and replaces System.Random
• Also replaces System.Security.Cryptography.RandomNumberGenerator (link)
• CryptoRandom is (unlike System.Random):
  • Fast (much faster than Random or RandomNumberGenerator)
  • Thread-safe (all APIs)
  • Cryptographically strong (seeded or unseeded)
• Implements Fast-Key-Erasure RNG for seeded CryptoRandom
• Produces the same sequence of seeded CryptoRandom values on all .NET versions (unlike Random)
• Wraps RandomNumberGenerator for unseeded (with additional smarts)
• Provides backtracking resistance: internal state cannot be used to recover the output
• Achieves ~1.3cpb (cycles-per-byte) performance, similar to AES-NI
• Scales per-CPU/Core
• Example: CryptoRandom.NextGuid() vs. Guid.NewGuid() [BenchmarkDotNet]:
  • 4~5x faster on Windows-x64
  • 5~30x faster on Linux-x64
  • 4~5x faster on Linux-ARM64 (AWS Graviton-2)
• Built for .NET 5.0+ and 6.0+
• Extensive test coverage & correctness validation (110+ tests)
  • CI runs on Linux-latest & Windows-latest

CryptoRandom API:

• All APIs of System.Random (link)
• CryptoRandom.Shared static shared instance for convenience (thread-safe of course)
• Seeded constructors:
  • CryptoRandom(ReadOnlySpan<byte> seedKey)
  • CryptoRandom(int Seed) (just like seeded Random ctor)
• byte[] NextBytes(int count)
• Guid NextGuid()
  • 5x (500%) faster than Guid.NewGuid() on Windows
  • 15x (1500%) faster than Guid.NewGuid() on Linux
  • 128 random bits, instead of 122
• Guid SqlServerGuid()
  • Returns new Guid well-suited to be used as a SQL-Server clustered key
  • Guid structure is [8 random bytes][8 bytes of SQL-Server-ordered DateTime.UtcNow]
  • Each Guid should be sequential within 100-nanoseconds UtcNow precision limits
  • 64-bit entropy for reasonable unguessability and protection against online brute-force attacks
  • ~15% faster than Guid.NewGuid()
• long NextInt64()
• long NextInt64(long maxValue)
• long NextInt64(long minValue, long maxValue)
• Random struct 's (make sure you know what you're doing):
  • void Next<T>(ref T @struct) where T : unmanaged
  • T Next<T>() where T : unmanaged

Utils API (SecurityDriven.Core.Utils):

• static Span<byte> AsSpan<T>(ref T @struct) where T : unmanaged
  • Casts unmanaged struct T as equivalent Span<byte>
• static ref T AsStruct<T>(Span<byte> span) where T : unmanaged
  • Casts Span<byte> as equivalent unmanaged struct T
• int StructSizer<T>.Size
  • Returns byte-size of struct T

Quick benchmark:

using SecurityDriven.Core;
using System;
using System.Diagnostics;
using System.Runtime.InteropServices;
using System.Threading.Tasks;

Console.WriteLine($"[{RuntimeInformation.FrameworkDescription}]");
Stopwatch sw1 = new(), sw2 = new();
const long ITER = 100_000_000, REPS = 5;

for (int i = 0; i++ < REPS;)
{
    sw1.Restart();
    Parallel.For(0, ITER, static i => CryptoRandom.Shared.NextGuid());
    sw1.Stop();
    Console.WriteLine($"{sw1.Elapsed} cryptoRandom.NextGuid()");

    sw2.Restart();
    Parallel.For(0, ITER, static i => Guid.NewGuid());
    sw2.Stop();

    var ratio = sw2.Elapsed / sw1.Elapsed;
    Console.WriteLine($"{sw2.Elapsed} Guid.NewGuid() [{ratio:N2}x slower]");
}

Output:
[.NET 6.0.0-rc.1.21451.13]
00:00:00.5486513 cryptoRandom.NextGuid()
00:00:02.0587652 Guid.NewGuid() [3.75x slower]
00:00:00.4117180 cryptoRandom.NextGuid()
00:00:02.0485556 Guid.NewGuid() [4.98x slower]
00:00:00.4103378 cryptoRandom.NextGuid()
00:00:02.0534771 Guid.NewGuid() [5.00x slower]
00:00:00.4100701 cryptoRandom.NextGuid()
00:00:02.0823213 Guid.NewGuid() [5.08x slower]
00:00:00.4017192 cryptoRandom.NextGuid()
00:00:02.0488105 Guid.NewGuid() [5.10x slower]

What's wrong with Random and RandomNumberGenerator?

• Random is slow and not thread-safe (fails miserably and silently on concurrent access)
• Random is incorrectly implemented:

Random r = new Random(); // new CryptoRandom();
const int mod = 2;
int[] hist = new int[mod];
for (int i = 0; i < 10000000; i++)
{
	int num = r.Next(0x55555555);
	int num2 = num % 2;
	++hist[num2];
}
for (int i = 0; i < mod; i++)
	Console.WriteLine($"{i}: {hist[i]}");
// Run this on .NET 5 or below. Surprised? Now change to CryptoRandom
// Fails on .NET 6 if you use seeded "new Random(seed)"

• Random/.NET 6 unseeded is fast (new algorithm), with a safe .Shared property, but instances are not thread-safe
• Random/.NET 6 seeded falls back to legacy slow non-thread-safe .NET algorithm
• Neither Random implementation aims for cryptographically-strong results
• RandomNumberGenerator can be much faster with intelligent wrapping and more useful Random API

Throughput (single-threaded) .NET 6:

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19042.1165 (20H2/October2020Update)
Intel Core i7-10510U CPU 1.80GHz, 1 CPU, 8 logical and 4 physical cores
.NET SDK=6.0.100-preview.7.21379.14
  [Host] : .NET 6.0.0 (6.0.21.37719), X64 RyuJIT

Throughput (single-threaded) .NET 5:

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19042.1165 (20H2/October2020Update)
Intel Core i7-10510U CPU 1.80GHz, 1 CPU, 8 logical and 4 physical cores
.NET SDK=6.0.100-preview.7.21379.14
  [Host] : .NET 5.0.9 (5.0.921.35908), X64 RyuJIT