The fastest storage for the fastest system: Summit

25 June, 2018
Douglas O'Flaherty
IBM

This week in Frankfurt, Summit was crowned the “Fastest System in the World” in the biannual Top500 survey. The ORNL system is the most powerful, with over 200 petaflops of NVIDIA and IBM Power Systems compute, but it is much more.

This is no ordinary system. It is purpose-built to use GPUs and excel on artificial intelligence (AI) workloads. AI differs from traditional high-performance computing (HPC) workloads. The majority of HPC is devoted to using computers to simulate the real world using mathematical models. There are many import applications, from drug development to weather predictions.

In contrast, AI programs are trained on historical data and predict new results as data comes in. In most cases, the data is unstructured, dense and diverse. These can consist of images, files, documents and datasets. Already a leader in data throughput, scalability and latency, IBM Spectrum Scale has been developed to deliver a file system for the fastest supercomputer in the world — one that was purpose-built for AI workloads with over 12,000 NVIDIA GPUs.

The aggregate numbers are staggering. Summit’s 250-petabyte storage system is delivered by a cluster of 77 IBM ESS Storage Systems that will deliver 2.5 TBs of data. But the real work has been on small file performance, where IBM Spectrum Scale has a 700 percent improvement in small file reads compared to our 4.2.2 version. Summit will have the capacity of 30B files and 30B directories and will be able create files at a rate of over 2.6 million I/O file operations per second. That is opening every book in the US Library of Congress in 10 seconds.

As impressive as these numbers are, it is really the impact these innovations can have for our clients, regardless of their size, that matters.  Whether it is the throughput required for a cluster of SAP HANA systems or GPUs, extreme low latency on NVMe all-flash, fast metadata search across billions of files, media serving or small file performance, the same IBM Spectrum Scale software that is driving Summit is available for you.

To read more about the Summit performance numbers read Ray Lucchessi’s analysis here.

The post The fastest storage for the fastest system: Summit appeared first on IBM IT Infrastructure Blog.