Dear all,
This post shares the HPC Solution of HPE.
2 Types of Products for HPC
Solution: Lustre + MSA low-end storage + D6020 high-density enclosure
Customer-oriented: Small- and mid-range customers, sensitive to costs
Customized solution: ClusterSTOR E1000 (integrated delivery)
Customer-oriented: High-end customers, mainly in North America
Ultimate bandwidth: 80 GB/s per 2 U
O&M tool: Cray ClusterStor data services
• While NFS remains the most widely adopted file system, it has dropped from being utilized at 54% of the sites down to 46%. NFS was one of the first file systems that could handle the initial scale of early HPC systems. Its first mover status, coupled with the fact that it’s still adequate in smaller scale HPC systems today, is reflected in its continued, albeit shrinking, wide adoption.
• Lustreutilization has grown from 21% to 32.5%, and Lustre’sopen-source approach has enabled it to mature its feature set across a wide number of areas including performance, resiliency, reliability, and scalability. This maturity has also given it stability and the capability of scaling to meet the demands of petascale and emerging exascale configurations.
• GPFS/Spectrum Scale adoption has grown at a slightly lower rate, from 23% to 26.8%. This modest growth in adoption can be attributed to a slower feature advancement pace due to its proprietary nature along with a market shift away from IBM’s dominance in the HPC sector and a change in the pricing model away from user-license-based to capacity-based.
This data confirms the selection of Luster as the parallel file system for the new HPC storage for the new era: Cray ClusterStorE1000 storage system.
Controller of CLUSTERSTOR E1000
Typical dual-controller design
Customized ODM for HPC
Optional NVMe SSD (1.92, 3.84, 7.68, and 15.36)
Disk enclosure of CLUSTERSTOR E1000
4 U 106 disks
5 U 84 disks
18 U provides a maximum of 5 PB available capacity.
Storage software: based on the open-source Lustre
Lustre enterprise feature providers:
GridRAID, T10-PI, pool quota, Scalable search, performance optimization
The performance of small files in Lustre 2.12 is optimized as follows:
Data on MDT (DoM) enables small files to be stored on the MDT (flash).
The MDT is configured with RAID 10 and dRAID to further improve performance.
Thanks.
