Drnglib

##Overview

drnglib (Digital Random Number Generator Library) is a Java 7 library that provides access to Intel Secure Key, the hardware random number generator introduced in the Ivy Bridge microarchitecture. This library retrieves cryptographically secure random values directly from the CPU using the rdrand instruction. Access to rdrand is implemented by a small native library embedded as a resource in the .jar file.

The DigitalRandom class is provided as a drop-in replacement for SecureRandom. It is thread-safe and stateless. All thread synchronization occurs in the hardware implementation of rdrand.

For more information see:

• Wikipedia: RdRand
• Intel DRNG Software Implementation Guide
• Analysis of Intel's Ivy Bridge Digital Random Number Generator PDF

##Usage

Follow these two steps to start generating random numbers:

1. Add the following dependency to your project (published on Maven Central):

    <dependency>
        <groupId>net.nullschool</groupId>
        <artifactId>drnglib</artifactId>
        <version>1.0.0</version>
    </dependency>

2. Create an instance of DigitalRandom located in the net.nullschool.util package:

    DigitalRandom random = new DigitalRandom();
    System.out.println(random.nextInt());

Your project will need Java 7 and an Intel Ivy Bridge (or newer) CPU. If the CPU does not contain a hardware random number generator, instantiation of DigitalRandom will throw UnsupportedOperationException.

drnglib runs on Windows, Mac OSX, and Linux. (Adding support for other operation systems simply requires building the native library code on the desired platform.)

##Performance

The following micro-benchmarks show the relative performance of the JRE's implementations of Random compared with DigitalRandom. These numbers were gathered using the benchmark harness contained in the test class PerformanceTest under the following conditions:

• Intel Core i7 3770S (quad core with hyper-threading enabled)
• Windows 7
• Java 1.7.0_11 64-bit
• -Xms1G -Xmx2G -XX:+PrintCompilation -XX:+PrintGCDetails
• Each thread provided its own RNG instance; no sharing.

Benchmark 1: Sustained generation rate of 64-bit longs (Random.nextLong)

MiB/sec Mean (Standard Deviation) over 20 trials after JVM warm up

+---------++-----------+-------------------++--------------+---------------+--------+
| Threads ||  Random   | ThreadLocalRandom || SecureRandom | DigitalRandom | Speed* |
+---------++-----------+-------------------++--------------+---------------+--------+
|    1    || 379   (0) |     2965   (0)    ||    40 (0)    |    192 (2)    |  4.8x  |
|    2    || 528 (263) |     5929   (0)    ||    75 (3)    |    382 (2)    |  5.0x  |
|    3    || 340  (69) |     8836  (32)    ||   100 (6)    |    569 (4)    |  5.7x  |
|    4    || 414 (118) |    11654  (78)    ||   129 (6)    |    747 (4)    |  5.8x  |
|    5    || 422  (29) |    14193 (103)    ||   141 (5)    |    760 (0)    |  5.4x  |
|    6    || 478  (37) |    16812 (462)    ||   147 (3)    |    760 (0)    |  5.2x  |
|    7    || 515  (29) |    19482 (812)    ||   158 (3)    |    760 (0)    |  4.8x  |
|    8    || 561  (43) |    21528 (911)    ||   167 (2)    |    758 (0)    |  4.5x  |
+---------++-----------+-------------------++--------------+---------------+--------+

*speed improvement of DigitalRandom over SecureRandom

Benchmark 2: Sustained fill rate of 2MiB arrays (Random.nextBytes)

MiB/sec Mean (Standard Deviation) over 20 trials after JVM warm up

+---------++-----------+-------------------++--------------+---------------+--------+
| Threads ||  Random   | ThreadLocalRandom || SecureRandom | DigitalRandom | Speed* |
+---------++-----------+-------------------++--------------+---------------+--------+
|    1    || 370   (0) |      644   (0)    ||    59  (0)   |    205 (2)    |  3.5x  |
|    2    || 652 (203) |     1287   (0)    ||   114  (6)   |    400 (3)    |  3.5x  |
|    3    || 841 (342) |     1791  (59)    ||   155 (11)   |    576 (8)    |  3.7x  |
|    4    || 910 (435) |     2092 (114)    ||   174 (11)   |    669 (6)    |  3.8x  |
|    5    || 935 (470) |     2224 (129)    ||   185 (11)   |    673 (7)    |  3.6x  |
|    6    || 750 (375) |     2324  (64)    ||   200  (8)   |    686 (8)    |  3.4x  |
|    7    || 854 (378) |     2460  (41)    ||   213  (7)   |    702 (6)    |  3.3x  |
|    8    || 694 (107) |     2508  (37)    ||   229  (6)   |    743 (8)    |  3.2x  |
+---------++-----------+-------------------++--------------+---------------+--------+

*speed improvement of DigitalRandom over SecureRandom

Benchmark 3: Sustained UUID generation rate

Millions UUID/sec Mean (Standard Deviation) over 20 trials after JVM warm up

+---------+-----------------+------------------------+---------+
| Threads | UUID.randomUUID | DigitalRandom.nextUUID |  Speed* |
+---------+-----------------+------------------------+---------+
|    1    |    3.4 (0.0)    |       12.4 (0.2)       |   3.6x  |
|    2    |    2.5 (0.1)    |       18.5 (3.4)       |   7.4x  |
|    3    |    2.4 (0.1)    |       26.8 (3.8)       |  11.2x  |
|    4    |    2.3 (0.0)    |       34.7 (3.7)       |  15.1x  |
|    5    |    2.3 (0.0)    |       41.8 (2.7)       |  18.2x  |
|    6    |    2.3 (0.0)    |       44.5 (1.3)       |  19.3x  |
|    7    |    2.3 (0.0)    |       45.7 (0.7)       |  19.9x  |
|    8    |    2.3 (0.0)    |       46.7 (0.7)       |  20.3x  |
+---------+-----------------+------------------------+---------+

*speed improvement of DigitalRandom over SecureRandom

Interpreting the results

DigitalRandom significantly outperforms SecureRandom, and even Random in some scenarios. SecureRandom on Windows defaults to the SHA1PRNG algorithm, but its implementation suffers from fairly trivial issues, such as the need to generate all values into byte[] buffers, even when a simple int or long is requested. Linux machines suffer further because SecureRandom's default algorithm gathers entropy from /dev/random, which can be very slow.

Java 7 clearly benefits from the introduction of ThreadLocalRandom, but it is important to note it does not produce randomness of cryptographic quality. It is an excellent replacement for Random, not SecureRandom.

The performance improvement DigitalRandom provides when generating UUIDs is particularly striking. This is because UUID.randomUUID(), which is the JRE garden path method for generating UUIDs, uses exactly one static instance of SecureRandom. Sharing one instance of SecureRandom causes severe thread contention, whereas DigitalRandom needs no thread synchronization.

##Build Requirements

###To build drnglib jar:

• Java 7 JDK
• Maven 3
• git

###To build drnglib's native libraries:

• gcc 4.2 or greater (Unix/Mac OSX)
• Visual Studio 2012 any flavor (Windows)

##Build Instructions

To build the Java jar:

1. get the source: git clone [email protected]:cambecc/drnglib.git
2. from the drnglib/ directory, do a maven build: mvn clean install

To build the native libraries:

NOTE: It is not necessary to build the native libraries unless you have made changes to the C code in the drnglib/src/main/c/ directory. Building the native libraries is somewhat tedious because each library flavor must be built from a toolset for the associated operating system. This usually means a Windows install for .dll, Mac OSX for .dylib, and Linux for .so. To build one particular flavor:

1. make sure the JAVA_HOME environment variable is defined to point to your Java 7 JDK
2. from the drnglib/src/main/c/ directory, invoke the appropriate build script, e.g., ./build-macosx.sh
3. rebuild the java jar to incorporate your native library. From the drnglib/ directory, do a maven build: mvn clean install

On Ubuntu, needed to get the following packages in order to build:

sudo apt-get install libc6-dev-i386
sudo apt-get install g++-multilib