> Just small additions like the following would help a lot: "[The X seconds] spent in fsync [seen in diagram A] accounts for almost exactly the delay we were seeing [in diagram B]".
Exactly. I try to always point to the actual data when I'm telling my readers my conclusions.
"Calls to futex, epoll_wait and select are largely the same on both systems. The problem host does 1727 calls to fsync(), while the normal host does only 849 calls. The problem host spends 1.24% of its time in fsync(). The normal host only 0.48%. This makes fsync() our number 1 suspect."
This at least lets the reader refer back to the output and see corresponding entries.
Exactly. I try to always point to the actual data when I'm telling my readers my conclusions.
"Calls to futex, epoll_wait and select are largely the same on both systems. The problem host does 1727 calls to fsync(), while the normal host does only 849 calls. The problem host spends 1.24% of its time in fsync(). The normal host only 0.48%. This makes fsync() our number 1 suspect."
This at least lets the reader refer back to the output and see corresponding entries.