go get github.com/awnumar/fastrand

fastrand implements a cryptographically secure pseudorandom number generator. The generator is seeded using the system's default entropy source, and thereafter produces random values via repeated hashing. As a result, fastrand can generate randomness much faster than crypto/rand, and generation cannot fail beyond a potential panic during init().

fastrand was forked from gitlab.com/NebulousLabs/fastrand. This version of the package uses securely allocated memory provided by github.com/awnumar/memguard.

fastrand also scales better than crypto/rand and math/rand when called in parallel. In fact, fastrand can even outperform math/rand when using enough threads.

Alternatively, use frand by Luke Champine if you are primarily concerned with speed or if memory security is not part of your threat model. frand is substantially faster than this package.

// 32 byte reads
BenchmarkRead32                     	10000000	       175 ns/op	 181.86 MB/s
BenchmarkReadCrypto32               	  500000	      2733 ns/op	  11.71 MB/s

// 512 kb reads
BenchmarkRead512kb                   	    1000	   1336217 ns/op	 383.17 MB/s
BenchmarkReadCrypto512kb             	      50	  33423693 ns/op	  15.32 MB/s

// 32 byte reads using 4 threads
BenchmarkRead4Threads32               	 3000000	       392 ns/op	 326.46 MB/s
BenchmarkReadCrypto4Threads32       	  200000	      7579 ns/op	  16.89 MB/s

// 512 kb reads using 4 threads
BenchmarkRead4Threads512kb           	    1000	   1899048 ns/op	1078.43 MB/s
BenchmarkReadCrypto4Threads512kb    	      20	  97423380 ns/op	  21.02 MB/s

fastrand uses an algorithm similar to Fortuna, which is the basis for the /dev/random device in FreeBSD. However, although the techniques used by fastrand are known to be secure, the specific implementation has not been reviewed by a security professional. Use with caution.

The general strategy is to use crypto/rand at init to get 32 bytes of strong entropy. From there, the entropy is concatenated to a counter and hashed repeatedly, providing 64 bytes of random output each time the counter is incremented. The counter is 16 bytes, which provides strong guarantees that a cycle will not be seen throughout the lifetime of the program.

The sync/atomic package is used to ensure that multiple threads calling fastrand concurrently are always guaranteed to end up with unique counters.