[Paper Review] Generating High Quality Random Numbers: A High Throughput Parallel Bitsliced Approach

Saleh Khalaj Monfared, Omid Hajihassani+4 more|arXiv (Cornell University)|Sep 10, 2019

Chaos-based Image/Signal Encryption33 references4 citations

TL;DR

This paper presents a high-throughput, cryptographically secure pseudo-random number generator (CSPRNG) using a parallel, bitsliced implementation on GPUs. By reorganizing data into column-major format and leveraging the LFSR-based MICKEY 2.0 stream cipher, the method replaces costly bit shifts with efficient register swaps, achieving 1.6 Tb/s on GTX 980 Ti and 2.9 Tb/s on Tesla V100—surpassing cuRAND in performance and cost efficiency while passing all NIST statistical randomness tests.

ABSTRACT

In this work, by employing a bitsliced data representation as building blocks of algorithms, we showcase the capability and scalability of our proposed method in a variety of PRNG methods in the category of block and stream ciphers. While demonstrating the suitability of stream-ciphers for high throughput PRNG, as an example, we implement and investigate a bitsliced MICKEY 2.0 PRNG by altering the paradigm of internal functions and data structure. The LFSR-based (Linear Feedback Shift Register) nature of the PRNG in our implementation perfectly suits the GPU's many-core structure due to its register oriented architecture and allows the usage of bit slicing technique to further improve the performance. In our SIMD vectorized fully parallel GPU implementation, each GPU thread is capable of generating a remarkable number of 32 pseudo-random bits in each LFSR clock cycle. We then compare our implementation with some of the most significant PRNGs that display a satisfactory performance in both throughput and randomness criteria. The proposed implementation successfully passes the NIST test for statistical randomness and bit-wise correlation criteria. To the best of authors' best knowledge, our method outperforms the current best implementations in the literature for computer-based PRNG and the optical solutions in terms of performance and performance per cost, while maintaining an acceptable measure of randomness. Our highest performance among all of the implemented CPRNGs with the proposed method is achieved by the MICKEY 2.0 algorithm which shows 1.9x improvement over the state of the art NVIDIA's proprietary high-performance PRNG, cuRAND library, achieving 1.6 Tb/s of throughput on the affordable NVIDIA GTX 980 Ti.

Motivation & Objective

• To address the performance bottleneck in high-throughput, cryptographically secure random number generation on general-purpose parallel platforms like GPUs.
• To improve upon existing PRNGs by enhancing throughput and cost efficiency while maintaining strong randomness guarantees.
• To adapt the bitslicing technique—previously used in symmetric ciphers like DES—for use in CSPRNGs, particularly LFSR-based stream ciphers.
• To demonstrate that column-major data representation enables full utilization of GPU SIMD datapaths, enabling highly parallel execution.
• To validate the method’s statistical quality using the NIST SP 800-22 randomness suite and compare performance against state-of-the-art solutions like cuRAND.

Proposed method

• The method employs a column-major data representation to enable full utilization of GPU SIMD datapaths, transforming row-wise operations into parallel bit-level computations.
• It applies the bitslicing technique to LFSR-based stream ciphers such as MICKEY 2.0, replacing expensive bit-shift and mask operations with efficient register swaps and bitwise operations (XOR, AND, OR).
• The implementation uses CUDA kernels with fixed thread block configurations (64 threads per block, 256 threads per grid) to maximize occupancy and coalesced memory access.
• Shared memory is utilized to reduce global memory latency and improve bandwidth efficiency during random number generation.
• The approach is generalized to other PRNGs, including AES and Grain, by adapting the bitsliced logic to their internal structures.
• Performance is normalized per processing power to enable fair comparison with prior works, especially on outdated or inaccessible GPU platforms.

Experimental results

Research questions

• RQ1 Can the bitslicing technique be effectively applied to CSPRNGs to achieve high throughput on GPU platforms?
• RQ2 Does column-major data representation significantly improve GPU utilization and performance in PRNGs compared to conventional row-major layouts?
• RQ3 How does the proposed method compare in throughput and randomness quality to established libraries like cuRAND and other high-performance PRNGs?
• RQ4 To what extent does the LFSR-based structure of stream ciphers like MICKEY 2.0 align with GPU’s many-core, register-oriented architecture for bitsliced execution?
• RQ5 Can the proposed method achieve both high performance and strong statistical randomness, as verified by the NIST SP 800-22 test suite?

Key findings

• The proposed bitsliced implementation on GPU achieves a peak throughput of 1.6 Tb/s using the MICKEY 2.0 algorithm on an NVIDIA GTX 980 Ti, representing a 1.9x improvement over NVIDIA’s cuRAND library.
• On the Tesla V100 GPU, the method achieves 2.90 Tb/s, significantly outperforming cuRAND while maintaining full compliance with NIST SP 800-22 statistical randomness tests.
• All 15 NIST statistical tests (including Frequency, BlockFrequency, CumulativeSums, and LinearComplexity) passed with p-values above 0.01, confirming the high quality of randomness in generated bitstreams.
• The method reduces costly bit-level operations (shifts and masks) by transforming them into efficient register-level swaps, enabling full parallelization across GPU threads.
• The performance per cost ratio is superior to both general-purpose PRNGs and optical solutions, making it a cost-effective high-throughput alternative.
• The approach demonstrates scalability across multiple GPU platforms, including GTX 980 Ti, GTX 1050 Ti, GTX 1080 Ti, and Tesla V100, with consistent performance gains over cuRAND.