AI AND HPC CONVERGED FOR YOUR STORAGE NEEDS
This confluence put new demands on HPC storage as the input/output patterns of both workloads could not be more different. And it is happening right now. A recent study of the independent analyst firm Intersect360 found out that 66% of the HPC users today already are running machine learning programs. Hyperion Research forecasts that, at current course and speed, HPC storage spending in public sector organizations and enterprises will grow 57% faster than spending for HPC compute for the next three years.
Seymour Cray once said, "Anyone can build a fast CPU. The trick is to build a fast system.” When it comes to HPC and AI, anyone can build fast file storage. The trick is to build a fast, but also cost-effective and scalable file storage system. We achieve this by embedding the leading parallel file systems into parallel storage products from HPE with cost effectiveness built in.
HPE STORAGE: BRINGING HPC AND AI TOGETHER FOR SCALED PERFORMANCE
| | Traditional HPC cluster | New Era Converged | Traditional AI stack (POD) |
|---|---|---|---|
Primary workload | Modeling & simulation | Both | Machine Learning/Deep Learning |
Compute node type
| CPU nodes like HPE Apollo 2000 | Both | GPU nodes like HPE Apollo 6500 |
Number of compute nodes | Hundreds to thousands | Hundreds to thousands | A few |
Typical interconnect | InfiniBand | Both (plus HPE Slingshot) | Gigabit Ethernet |
Primary I/O pattern | Write-intensive | Both | Read-intensive |
Storage capacity measured in | Petabytes | Petabytes | Terabytes |
Storage scalability in single namespace | Up to exabytes | Up to exabytes | Up to low double digits petabytes |
Typical storage | HDD-based parallel file system storage | Cray ClusterStor E1000 HPE Parallel File System Storage | All Flash Enterprise NAS storage |
Storage well suited for | Serving large files in sequential order at speeds of up to terabytes per second | Serving files of all sizes in both sequential and random order at speeds of up to terabytes per second | Serving files of all sizes in both sequential and random order at speeds of up to double digit gigabytes per second |
Price per terabyte | $ | $$ | $$$$$ |
SCALABLE, COST-EFFECTIVE AND CONVERGED PERFORMANCE
Designed to deliver performance and scalability far beyond scale-out NAS, our storage systems use leading parallel file systems, and do not require capacity-based licensing of the file system. HPE Pointnext provides support for the complete storage system, hardware and software.
| | HPE Parallel File System Storage | Cray ClusterStor E1000 Storage System |
|---|---|---|
Parallel file system | IBM Spectrum Scale Erasure Code Edition | Lustre |
Typically attached to | Clusters of HPE Apollo Systems or HPE ProLiant DL rack servers with HPE IB HDR/Ethernet 200Gb adapters | HPE Cray EX supercomputers or very large clusters of HPE Apollo systems and HPE ProLiant DL servers |
Number of compute nodes | 10s to a few 100s | Several 100s to 1,000s |
Type of compute nodes | GPU and/or CPU | GPU and/or CPU |
Connected with | InfiniBand HDR or 100/200 GbE | HPE Slingshot or InfiniBand HDR or 100/200 GbE |
Set of enterprise grade storage features | Comprehensive | Basic |
Protocol support | POSIX, NFS, SMB, HDFS, Object, NVIDIA GPUDirect Storage (GDS) | POSIX, NVIDIA GPUDirect Storage (GDS) |
Scale-out typically in | Terabytes (starts at 27 TB) | Petabytes (starts at one PB) |
Best suited for HPC and AI in | Manufacturing, healthcare and life sciences, financial services, high tech, pharma, telco, retail, and for home directories in organizations listed in the right column | Research laboratories and universities, defense and intelligence agencies, space agencies, climate research, weather forecasting, and geosciences, including oil and gas |
Parallel storage solutions for the new era
Whether you prefer IBM Spectrum or Lustre, we have the right parallel storage system for you. And with HPE Data Management Framework (DMF) V7 you can move files between them or out of them for data protection or long-term archive purposes.
This storage system embeds Lustre, the leading parallel file system for scientific and federal use cases. It’s ideal for attachment with HPE Slingshot, InfiniBand HDR and 100/200 GbE to HPC Cray EX supercomputers and large clusters of HPE Apollo systems or HPE ProLiant DL rack servers running modelling and simulation, AI or high performance data analytics workloads. HPE Pointnext Services support both hardware and software.
This storage system embeds IBM Spectrum Scale, the leading parallel file system for enterprises. It’s ideal for attachment with InfiniBand HDR and 100/200 GbE to clusters of HPE Apollo systems or HPE ProLiant DL rack servers running modelling and simulation, AI or high performance data analytics workloads. HPE Pointnext Services support both hardware and software. Storage servers are cost-effective HPE ProLiant DL rack servers.
Optimize HPC storage resources by enabling a hierarchical, tiered storage management architecture. The HPE Data Management Framework optimizes storage resource utilization for lower TCO and streamlines data workflows with time proven automated data movement across the full storage hierarchy. Now with native support for IBM Spectrum Scale in addition to Lustre and HPE XFS.