Our newest storage innovations tailor HPC storage to specific project needs
Cray® ClusterStor™ L300N Storage with NXD
Storage access time has been identified by Hyperion as the #2 roadblock in HPC. The ClusterStor™ L300N hybrid storage system with intelligent NXD software is purpose-engineered to address the mixed I/O bottleneck that many HPC users are facing as they run more and more applications on the same system.
The non-distruptive acceleration of data (NXD) software stack contains features that address the diverse needs of modern HPC applications:
- Read Persistence: Uses advanced caching technologies to enhance small block random I/O performance by identifying frequently accessed data.
- Write Back: Allows writes at full speed to the NXD storage tier, gathers the datasets, then writes to disk over a period of time in the background.
- I/O Histogram: Provides the profiling of specific application workloads from a storage point of view. Understanding where data ultimately lands on the storage is key to setting the correct tunable parameters for NXD.
- Performance Statistics: Generates real-time updates of the cache device performance — such as cache hits, cache usage level and cache space consumption — essential for optimal performance.
- Dynamic Flush: Flushes of dirty data are performed by the flush manager of NXD under different scenarios, such as when the amount of dirty data reaches a certain threshold or when I/O activity drops.
Cray® View for ClusterStor™
The Cray® View for ClusterStor™ storage system monitoring application enables customers to correlate performance of ClusterStor storage systems to jobs running on Cray XC series supercomputers. Cray View for ClusterStor technology provides visualization of performance data for overall storage systems and individual jobs, including throughput and metadata activity. Users can define, detect, and be alerted about poorly performing jobs that may be aggressively using system resources and impacting other running jobs.
Cray® DataWarp™ Accelerator with Transparent Caching
Not only is DataWarp I/O acceleration technology now available for the XC50 supercomputer in two- or four-node configurations, it now includes transparent caching.
In this mode, the DataWarp accelerator operates as a file-level cache of the parallel file system, which means:
- Applications see the directories of the parallel file system, not a private file system
- DataWarp nodes will read-ahead files being read
- Write-behind of files being written
- Several flush options: close, fsync, as needed, and at the end of the job
- Application knobs to flush, invalidate, etc.
- With this mode XC users can now use (all-flash) DataWarp for data access performance and (all-HDD) ClusterStor L300 for storage capacity
Unlike other burst buffer approaches, DataWarp technology leverages all the proven performance advantages of the XC series’ Dragonfly network topology through direct-attach to the Aries™ interconnect. In addition, it enables a broad set of HPC tools — from workload managers to job schedulers to user scripts — to increase the productivity of application processing using in-system I/O acceleration.