I have written a small program that is artificially (and sadly also quite stupidly)
aging an empty file system. Basically it works in the following way: It calculates the percentage of
used blocks fsb, which is the number of used blocks in the file system divided by the number of
blocks the partition has. It determines the probability X that it creates (or with
probability of 1-X deletes)
a file (given certain distributions of the file size). The create/delete probability tableX is given by
| X |
between fsb |
| 1.00 |
0-2 % |
| 0.90 |
2-30 % |
| 0.85 |
30-50 % |
| 0.80 |
50-85 % |
| 0.70 |
85-90 % |
| 0.60 |
90-94 % |
| 0.50 |
94-97 % |
| 0.30 |
97-98 % |
| 0.00 |
98-100 % |
The program measures the I/O (currently only write) performance in counting bytes written divided through time needed in 5 percent steps of used blocks.
The directory structure can be chosen to be quite flat (at the moment just two hierachies, I use
100 times 50 directories), which is fair enough for all different file systems including ext2 (that is less
than 30 files per directory with the chosen distribution of the file size for a partition of 500MB).
I plan to extend the program that it not only creates new files, but also appends to existing
files. In a further step it should also measure read performance (which I am ignoring at the
moment).
Later on I will measure how the performance behaves with time: when reaching less 5-10 percent
of free blocks, let the program just invert the above table for the probability until reaching 90-95
percent of free blocks, then starting again with the normal table. If fragmentation effects show of, the
performance should become less with each cycle which one can show by dividing the result of each cycle-turn through
the result of the first cycle ("gauge") measurement.
