HP-UX Encrypted Volume and File System Performance and Tuning

EVFS Testing With Postmark
All testing so far has been done with the IOZone benchmark tool, which is very effective for profiling
input/output system characteristics. However, IOZone does not represent most user applications. The
Postmark benchmark was created to more accurately represent actual user applications.
Postmark can create very large file sets, and then operate on existing files after their creation. The
application profile was intended to resemble a mail server. The performance measures are in
throughput via reads and writes (like IOZone) and also in transactions per second.
For the EVFS test cases, 100,000 files were created in a single subdirectory, in sizes ranging from
10KB to 1MB. The read/write size was 8kb (like most applications).
PostMark 8k Block - 10k-1mb Files, 100k Tranactions
9%
12%
23%
27%
JFS 4.1 JFS 5.0
HP-UX File System Version
CPU Utilization %
Throughput KBs
Clear APP CPU EVFS APP CPU
Clear Read EVFS Read Clear Write EVFS Write
Figure 28: Postmark Testing
This graph shows the differential between clear I/O and EVFS using Postmark, and also the
differential between a default system with JFS 4.1 (the default file system) and JFS 5.0. JFS 5.0
includes enhancements to more efficiently handle large numbers of files.
Using JFS 5.0 and simple VxFS tunes increased both EVFS I/O and clear I/O by 28% for reads (the
lower pair of lines on the graph) and 22% for writes (the upper pair of lines). These increases
required additional CPU – 3% more for clear I/O and 4 % more for EVFS.
Notice that the read throughput results for EVFS are quite close to clear I/O – 81%. Although the
write throughput results for EVFS appear to have a higher spread, the actual results for EVFS writes
are 87% of clear I/O. In this case EVFS actually performed quite closely to clear I/O for a more
representative benchmark of an actual application.
EVFS used a little more than double the CPU of clear I/O to achieve these throughput results.