White Papers

Dell Storage for HPC with Intel Enterprise Edition for Lustre sofware
Figure 12: File Metadata Operations
Figure 12 illustrates the file metadata results using MDtest. From this graph, file create metadata
operations start with less than 504 OPS at 1 thread and scale to almost 15K OPS with 240 concurrent
threads. This may be due to the Lustre locks needed on the MDT, but also those on the OSTs, since
using a stripe count of 24 had a significant decrease in performance. At 240 threads, we had 2
directories (-b 2) and 2.2 million files were created.
File stat metadata operations is overwhelmingly the lightest metadata operation of the three observed.
A single thread test yield over 2K OPS and scale to more than 400K OPS with 240 concurrent threads.
The increase in performance could be due to improvements made in Lustre version 2.5 metadata
operations. Also of note is the use of 15K RPM drives in the MDT volumes.
Removal of files is also limited by accesses to OSTs, similar to the create operation. However, the
remove operation have advantage over create operations when increasing in total threads, starting
with over 1.8K OPS at 1 thread and scaling to almost 100K OPS with 192 concurrent threads.
Figure 13 illustrates the directory metadata results using MDtest. From this graph, directory create
metadata operations is again the most expensive operation for most cases, starting with 1.8K OPS at 1
thread reaching a maximum of almost 19K at 128 threads. Directory operations are also affected by
the number of top directories used (-b), but to a lesser degree than file operations. Remove operations
is almost as expensive as directory createsas it starts with almost 1.9K OPS at 1 thread and reaches a
maximum of over 70K OPS at 120 threads. Directory Stats operation is again overwhelmingly the
lightest of the three observed. A single thread test yield over 5K OPS at 1 thread and grows to over
450K at 128 threads.