performance
Warning: This is an ancient artifact.
The tycho cluster at LLNL includes a gigabit ethernet-attached Lustre file system, an infiniband-attached compute cluster, and LNET routers between the two. The design, which mimics the storage architecture in LLNL's production computing environment, allows Lustre to be directly mounted on the compute nodes. In this experiment, Lustre is instead mounted on one of the LNET routers, and diod is used to remount Lustre on 64 compute nodes. The two approaches are then compared with I/O benchmarks.
The file system is capable of around 5 GB/s, but in this case one router is used so the network limits I/O bandwidth to around 2.5 GB/s.
IOR 2.10.2 was run chdired into a unique test directory in Lustre, with the following command line:
ior -i 16 -a POSIX -w -r -t 2m -b 1g -W -C -e -F -g -v -d 1 -o testfile
This is the ior read performance:
This is the ior write performance:
