Hi fellow programmers/developers,
I am trying to measure xfs performance on our raid storage. While doing that, I see an unusual observations. Before going to the observations, few details: 1. I know that while writing to xfs repetitively with each iteration in a different file, we see a saw - tooth pattern. The write performance in GB/s starts from a certain point say 6 GB/s and degrades to a point where it is around 4.2 GB/s and jumps back to 6 GB/s in next iteration. The cycle repeats after a fixed number which depends on the allocation blocks in the storage. I understand it's because we are writing in allocations blocks starting from outermost disk cylinders towards innermost cylinder on the disks. I am sorry I am a new user and could not attach the saw-tooth pattern picture. The observations I made is that this saw tooth pattern only appears when I write the repetitions of the test in different directories. If I write the repetition in directly under a single folder, there is no saw tooth pattern. For example :
Main Directory
------repetition1.dat
------repitition2.dat
------repitition2.dat
does not shows saw tooth pattern shown in attached image, while
Main Directory
----Directory1
---Repetition.data
----Directory2
---Repetition.data
----Directory3
---Repetition.data
does show a saw tooth pattern.
Can someone explain why this happens.
Other data points
We are using CentOS 7.0
There is no effect of choosing a IO scheduler. The behaviour is same with all IO schedulers
