File and Object Storage

IBM Spectrum Scale for AI workloads beats the competition with hybrid cloud Global Data Services

By David Wohlford posted Mon March 08, 2021 02:45 AM

  

Edited by Peter Basmajian

Performance is a key part of AI, analytics and HPC workloads.  The problem with only measuring performance is that requirements continue to change.  For example, although supercomputers from a few years ago still offer high performance, they are being replaced because they are not as fast as today’s more efficient systems and they are very expensive to maintain.  AI, analytics and HPC data requirements are similar to processing performance requirements and will continue to grow each year.   With this growth in data requirements comes more demands on performance for the data.  Storage vendors will continue to leapfrog each other in performance and throughput numbers in the coming years -- but is that all that really matters?   

Barriers to Implementation

Gartner recently released a report on barriers to AI implementation. (https://www.gartner.com/smarterwithgartner/build-3-operations-management-skills-for-ai-success/) It’s interesting to note that data performance was not listed as a barrier to AI implementation.   Data volume and/or complexity, data quality and data accessibility were the top data barriers. Performance matters for AI, but it’s expected that vendors must provide it.  Software is the differentiator that can provide long term investment protection and data agility as well as flexibility for changing requirements. 



IBM Spectrum Scale has provided high performance data solutions for almost 30 years, has over 200 technology patents, and is built upon an adaptable architecture that continues to evolve with  modern workloads.  From its inception, the product was designed to support high throughput workloads and applications.  Spectrum Scale is deployable on commodity X86 servers, Elastic Storage System (ESS) building blocks, Kubernetes containers and Red Hat OpenShift, including on multiple public clouds such as AWS and Azure.  IBM provides global cloud data services from edge, to core, to cloud.

As part of a recently published reference architecture with IBM Spectrum Scale for the NVIDIA A100 using the IBM ESS 3000 ( www.ibm.com/downloads/cas/MJLMALGL ), IBM was able to demonstrate performance throughput of 94GB/s in a 2-node linear scaling configuration.  This performance for a parallel file system like IBM Spectrum Scale is considerably ahead of NFS scale-out node performance benchmarks, and is on par with similar storage solutions with NVIDIA. 

IBM continues to innovate for high performance workloads, and you should continue to see performance innovation throughout 2021.  Benchmarks are only one indication of actual performance.  Some can be tuned to achieve a specific number or workload, but may not represent actual AI workloads that simultaneously access small, medium and large files on separate concurrent applications. 

Global Data Services

But now let’s look at the software and the value for data and not just performance.  Optimum performance can only be achieved if the data is available for the workloads being measured. The cost of storing data must also be optimized, or datasets will be limited by the amount of high-performance data, and not as useful for AI analysis.  Data must also be secure and provide protection even from outside or inside attacks.

Only IBM provides Spectrum Scale and ESS with global cloud data services.  IBM data services are proven and comprehensive, including thousands of customers across multiple industries and use cases.  Spectrum Scale can access data on modern Kubernetes platforms, legacy storage systems, object storage systems, or even on the public cloud. Accessing this data without complex data movement or large delays can take more time than actually running AI workloads.  Data must be manageable, or data silos will eventually create complexity that will limit AI business results.

 

When it comes to data agility and enterprise data access, IBM Spectrum Scale is hard to beat.  Customers can create a single 8 YB namespace which spans everything from object or legacy data to concurrent data in the public cloud or within HDFS applications.  It’s the software with integrated simplicity that propels IBM ahead of the competition.

For more information on Spectrum Scale click here

For more information on ESS click here

Download the latest IBM Spectrum Scale and ESS NVIDIA A100 reference architecture

 

 


#Highlights-home
#Highlights
0 comments
1076 views

Permalink