PowerScale P100 performance accelerator
PowerScale P100 performance accelerator
-
The P100 performance accelerator node is added to the PowerScale cluster by Ethernet or InfiniBand connections to the back-end switches. Each P100 node provides additional compute to the cluster and front-end ports for added connectivity and network bandwidth directly to the compute clients.
Figure 3. PowerScale P100 performance accelerator connectivity
During internal lab testing of a PowerScale cluster with four F900 nodes and one P100 performance accelerator node, we saw an increase of approximately 13 percent in streaming NFS read throughput.
Note: Actual performance results will depend on customer workflow profile, cluster configuration, and load.
Note: Deploying one or more P100 performance accelerator nodes with PowerScale F900 and/or F600 all-NVMe flash nodes is advised. Although the P100 can also boost performance when used with other PowerScale node types, performance depends on the actual workflow and cluster configuration.
The P100 node should not be considered as a general solution for all types of workflows or use cases. The P100 accelerator is most suited to serialized workflows where it can provide one or more of the following benefits:
- Magnify cached read performance and reduce latency on read-intensive workflows
- Increase high single-stream throughput
- Allocate incremental performance for individual workloads
- Expand front-end connectivity for fan-in and aggregate network bandwidth improvement
- Provide additional performance for clusters that do not require additional capacity
The P100 does not provide significant performance improvements in workflows that include any of the following characteristics:
- High levels of jobs running concurrently, which induce large amounts of metadata overhead (such as EDA benchmark workflow)
- Uncached concurrent reads
- Streaming, concurrent writes
Instead, with these types of workflows, consider additional storage nodes to improve performance.