Conclusions
OK, let's summarize:
- “near” layout is a very reasonable default, as it give very good random I/O and passable sequential speed;
- “offset” layout is my favorite: it had good random I/O and good sequential speed, showing the better of the two worlds;
- “far” layout has very high sequential speed, but its lower random I/O performance means that, unless you are sure that your workload is mainly composed of sequential operations, you should avoid it.
Remember that the above tests were done on a 100 GiB disk partition (~200 GiB useful array size). This means that, will full disk partitions, “far” random I/O results can be significantly lower then expected.
Please also note that, while IOMeter is a wonderful tool, it had its limitation. Moreover, no synthetical test can cover 100% of real-world scenarios so it would be preferable to do some quick testing of your specific needs. In my case, as I will use this array to host virtual machine files, I found IOMeter to accurately represent my usage pattern (low metadata overhead, many read/write operations on one or few files, etc.) and so I happily used it ;)
I hope that you find this article interesting. If you want, you can discuss it with me using the comment system or by writing at
Have a nice day!