Is your scientific software complemented by the best hardware?
Did you know you could realize huge performance gains for your scientific applications by taking advantage of the unique capabilities of the IBM Power Systems platform? Are you apprehensive about making the leap to another architecture? Let’s look at the current state of the software ecosystem around IBM POWER, the resources available and how IBM Systems Lab Services can help you take that first jump to success.
The ever-growing interest in Power Systems for deep learning and other scientific applications
IBM Power Systems are increasingly being adopted as an alternative to x86-based systems. The advanced capabilities of the Power platform (wider cache lines, increased memory bandwidth, NVLink, SMT, AI capabilities and so forth) are especially appealing to scientific communities. These characteristics bring unprecedented performance and productivity benefits for large-scale research and development activities (numerical simulations, deep learning tasks, big data analytics and so forth).
The recent adoption of the Power Architecture for the Summit and Sierra supercomputers (respectively the first and second fastest in the world today), and the deployment of IBM Power Systems in leading academic institutions (Oregon State University, New York University, the University of Birmingham and others) indicates the continued growing interest in the Power platform for large-scale deep learning and other scientific workloads. In parallel, the software ecosystem around the Power platform is steadily growing and evolving, with a strong network of software providers benefiting from the Power Architecture.
Open source matters
The scientific community, to a large extent, relies on open source software. Enterprises are also increasingly adopting open source software in their production environments, as evidenced by the recent surge of open source deep learning frameworks and the ubiquity of their adoption in both academia and industry.
In tune with the open source community, IBM and the wider OpenPOWER ecosystem have developed pre-built binaries targeting the Power Systems architecture for a range of popular software in key scientific areas. The Open Source POWER Availability Tool (aka OSPAT) is a good way to check where to find the packages you need. The freely available IBM PowerAI software provides easy access to an environment with popular open source deep learning frameworks (Tensorflow, Caffe, PyTorch and so on) optimized for the Power Architecture, helping you to get your environment up and running quickly. The availability of Anaconda distribution on POWER greatly eases the installation and maintenance of several data science (such as pandas, scikit-learn, H2O, xgboost) and other python packages.
Porting your software to Linux on Power
With POWER8- and POWER9-based systems supporting little endian, porting software from other architectures to Power Systems is becoming a breeze. The how-to porting guide is the best place to make a start, with details of the steps to take and additional resources to help in the process. Often, porting software onto the latest Power processor system simply involves compiling the source code, executing and verifying the correctness of installation using sample problems. However, coding practices heavily focused on x86 optimizations can sometimes cause subtle performance issues. It’s useful to be aware of techniques for mitigation of any such performance issues. For scientific software that needs nontrivial work to build on Power Systems, a number of IBM tools and approaches exist to aid in the process.
Immediate performance benefits
So, is it worth the effort of porting your software to Power Systems? Order of magnitude speedups have been demonstrated for popular scientific software in various fields, including computational chemistry, bioinformatics, high energy physics and artificial intelligence. Now your software could also run much faster by taking advantage of Power-based systems. This huge increase in turnaround times can drive your productivity as well as cost savings.
To help organizations moving to Power, IBM Systems Lab Services consultants offer porting and tuning services. We can also enable your staff to perform porting through skills transfer. Contact Lab Services today for porting and tuning services.