Unlocking the Power of Knights Ferry Cards in High-Performance Computing

Table of Contents

1. Introduction
2. Overview of the System
3. Initial Experiences with the Knights Ferry Cards
4. Programming Experiences with Knights Ferry
5. Performance Results on Knights Ferry
6. Scalability and Affinity Settings on Knights Ferry
7. Porting Applications to Knights Ferry
8. Performance Comparison with Other Libraries
9. Finite Element Code and Performance Results
10. Lattice Boltzmann Algorithm and Performance Results
11. Conclusion

Introduction

In this article, we will explore the early experiences and programming efforts on the Knights Ferry cards. These cards are part of the Stampede system, which is a 10 petaflop supercomputer built by TACC (Texas Advanced Computing Center). We will discuss the system specifications, the challenges faced during porting applications to Knights Ferry, and the performance results obtained from various benchmarks and algorithms.

Overview of the System

The Stampede system is a powerful supercomputer that combines traditional Sandy Bridge CPU clusters with the Knights Ferry cards, which are specialized coprocessors for high-performance computing. The system features over 6000 nodes, totaling more than 100,000 cores of Sandy Bridge CPUs and Knights Ferry coprocessors. Additionally, the system boasts a total of 14 petabytes of disk storage and 200 terabytes of RAM. With its large-Scale infrastructure and advanced capabilities, Stampede is poised to become one of the largest and most powerful supercomputers ever built.

Initial Experiences with the Knights Ferry Cards

When working with the Knights Ferry cards, the initial experiences were highly positive. The codes were relatively easy to port, and most of the effort involved rewriting build scripts and makefiles. The system showed perfect scaling with OpenMP, thanks to the large L3 cache that could hold significant amounts of data. The Knights Ferry cards proved to be capable coprocessors, delivering exceptional performance and strong scalability on Parallel applications.

Programming Experiences with Knights Ferry

The programming experiences with Knights Ferry showcased its potential for high-performance computing. The Knighs Ferry cards proved to be well-suited for parallel programming, and the porting of applications was straightforward. Most codes compiled without issues, and the performance improvements were remarkable. However, achieving optimal performance on Knights Ferry required additional effort, such as optimizing vectorization and exploring different affinity settings.

Performance Results on Knights Ferry

The performance results obtained on Knights Ferry were highly promising. The system showed excellent scalability, with the applications demonstrating strong speedup as more cores were used. The inclusion of the Knights Ferry coprocessors significantly enhanced the overall performance of the system. Additionally, benchmarks and tests indicated that Knights Ferry outperformed other libraries, such as FFTW, thanks to its improved vectorization and optimized implementations.

Scalability and Affinity Settings on Knights Ferry

Scalability was a key aspect of the Knights Ferry cards, and it was crucial to explore different affinity settings to achieve optimal performance. Tests conducted on various applications revealed that the scatter affinity setting provided the best overall performance, especially at larger thread counts. Furthermore, it was found that native compilation and the use of static libraries improved the robustness and ease of use of the Knights Ferry cards.

Porting Applications to Knights Ferry

Porting applications to Knights Ferry was a relatively straightforward process. The majority of the effort involved rewriting build scripts and makefiles to ensure compatibility with the system. Native compilation and the use of static libraries were sufficient for most applications. However, it was important to test and optimize the applications for the Knights Ferry architecture to achieve the best performance and scalability.

Performance Comparison with Other Libraries

The performance of Knights Ferry was compared with other libraries commonly used in high-performance computing. It was found that the Intel Math Kernel Library (MKL) consistently outperformed libraries like FFTW in terms of speed and scalability. The improved vectorization capabilities of MKL contributed to its superior performance on Knights Ferry. Additionally, affinity settings played a significant role in optimizing the performance of the applications.

Finite Element Code and Performance Results

One of the applications tested on Knights Ferry was a finite element code. This code, primarily MPI-based, demonstrated excellent scaling and performance on the Knights Ferry cards. It was found that the Knights Ferry architecture benefited from vectorization and openMP parallelization of certain loops. However, further optimization and offloading strategies were required to fully exploit the capabilities of the Knights Ferry cards.

Lattice Boltzmann Algorithm and Performance Results

Another algorithm evaluated on Knights Ferry was the lattice Boltzmann method. Extensive vectorization and optimization efforts were made to improve the performance of this algorithm. The results showed a significant speedup on both Knights Ferry and Westmere architectures. However, the scalability of the lattice Boltzmann algorithm on Knights Ferry reached a plateau beyond a certain number of Threads, highlighting the need for further optimization and parallelization strategies.

Conclusion

The early experiences and performance results on the Knights Ferry cards showcased their potential for high-performance computing. The porting of applications to Knights Ferry was relatively straightforward, and the system demonstrated strong scalability and impressive performance on parallel codes. The comparison with other libraries and the exploration of affinity settings provided valuable insights into optimizing the performance of Knights Ferry applications. With ongoing advancements and improvements in silicon technology, Knights Ferry continues to prove its capabilities as a powerful coprocessor for scientific and computational research.

🌐 Resources:

• Texas Advanced Computing Center (TACC)
• Intel Math Kernel Library (MKL)
• FFTW