By Henk Vandenbergh-Oracle on Feb 11, 2010
I have written a blog entry about problems with the patterns= parameter before, even mentioning that I may no longer support it. I have concluded since, that I need to continue supporting it though in a different format than currently, where you (in older versions) could specify 127 different data patterns.
In Vdbench 5.01 and and 5.02 (brandnew), patterns= works as follows: patterns=/pattern/dir where file name '/pattern/dir/default' gets picked up, and its contents stored in the data buffer used for writing data. That works.
However, (and these things always happen when it is too late) a few hours after I did the last build of Vdbench 5.02 I realized that yes, I put the pattern in the buffer, but I use the same buffer for reading which means that if you have a mixed read/write workload your data pattern can be overlaid by whatever data is on your disk. Since the pattern is copied only once into your buffer all new writes will NOT contain this pattern. So, until I fix this, if you want a fixed pattern to be written, do not include reads in your test.
In normal operations I use only a single data buffer, both for reads and writes. This is done to save on the amount of memory needed during the run. Loads of luns \* Loads of threads = Loads of memory. This now needs to change when using specific data patterns.