Random read/write speed
While traditional, rotating disks are quite OK at sequential operations, random I/O is a very different beast: not considering caches and I/O coalescing, a disk with a 10 ms access time will at most complete 100 read/write operations per second. At ~4KiB per operation, this add up at only 400 KiB/sec.
Sure caching and other techniques can improve this rating, but the reality is that platter-based disks remain slow at random operation (for example, a good SSD can be 20x faster then the fastest mechanical disk).
This, coupled with the fact that most server workloads are random in nature, means that a good RAID had to give back a reasonable speedup over single-disk operations, otherwise it will be not so useful.
Let's start with read speed:
All layouts provide very similar results here, but the “near” layout is the leader.
“Near” and “offset” layouts are on par, while “far” is a bit behind. However, it is somewhat surprising to see that “far” layout is not so slow: given the additional seeks it imposes, I expected it to be even slower. Anyway, we had to remember that I am testing with 100 GiB partitions: this means that when using the full disk space (~900 GiB), it can be noticeably slower (as the disk heads need to travel additional distance for each seek).
Last, mixed (50% read – 50% write) speed:
Here we see the “near” layout on top, closely followed by “offset”. On the other hand, “far” is significantly slower (remember that this can only get worse if we use full disk space).