My machine has 4 SATA 2 West Digital 1TB disks. I made 2 partitions on each of them, 500GB for each partition. When I started using them I check their I/O using iozone. The first partition has 100MB/s for read, 70MB/s for write. And the second partition has 80MB/s for read, 55MB/s for write. All 4 disks has the same result.

As I use on, the I/O speed on each partition decrease, to different extend. For example, for the 4 first partitions, the write speed varies from 69MB/s to 56MB/s. And I have same amount of data on each of them, all used 11%.

My guess for this is the disk block allocation policy. This is caused because some disk starts writing from inner location while others writes on the outer edge, even though amount of data on each disk is the same.

Any expert idea for this ?

Well. "Same amount of data" does not equal to "same data" on the 4 disks. If these disk were all mirrors of each other, then the I/O would probably be identical. If as I understand from you, each disk is written to independently, then the difference is expected. Every write operation is evaluated by the filesystem & written to different areas of the disk.

Besides that is the factor that each disk writes in a different way (from inner to outer sectors in the platter or the other way around) and the fact the the data could well span multiple platters.

The first sectors of the disk are on the outer tracks, where there are more bits per physical track, so more bits per second in transfer rate as well as more bits reachable within a given physical seek distance.

Partitions get fragmented. Many documents and posts will tell you that fragmentation doesn't matter in Linux file systems. That isn't really true. Fragmentation doesn't matter as much in sane filesystems as it did in FAT. Fragmentation doesn't matter as much in the way files are used in modern systems as it did many years ago. But fragmentation will still produce measurable differences.

Why would the policy be different?

It seems more likely that the sequence of creation, extension and deletion of files varied among your partitions. Even if the total fraction used ended up the same, the exact sequence of creation, extension and deletion requests that reached that current use level could result in arbitrary differences in the distribution of the used sectors.

I think all disks give the lowest physical addresses to the outer edge and the highest physical addresses to the inner edge.

I think the file system policies will tend to use lower physical addresses before higher physical addresses, but that will just be a tendency, not a strict rule. Lots of factors may influence which sectors a filesystem decides to use next.