在更新的Linux内核上，同步IO的速度令人难以置信

Question

I just find that sync IO is getting much much slower on newer kernels. I have tried the following commands on CentoOs(2.6), Fedora20(3.11), Ubuntu12(3.2) .

echo 0 > test
dd if=/dev/zero of=test bs=1024 count=200 oflag=dsync

The speed for 2.6 is about 1MB/s and only around 25kB/s for new kernels.

I have a CFD code which calls sync() in a C function. Usually the code runs very fast on old kernels, but now it is much slower on newer OS. I am really curious what the Linux developers have done to sync() . The other possible reason is not on kernel version, but the file system used by different OS.

At this point my code needs sync() anyway, so is there anything I can tune like ioctl() to speed it up?

Any comment is truly appreciated.

Answer 1

The apparent performance of sync() is greatly affected by the amount of memory installed in the system, the usage pattern of disk i/o (including the drive's caches, DMA controllers, etc.), and the speed of the disk drives.

Without rigorously controlling for all these variables, comparing one system to another is reasonably meaningless when you think you are changing only the kernel.

The effect of newer hardware seeming to make things slower makes sense. Instead of the kernel flushing out pending writes when there is a 250 MB backlog (because the system only has 2 GB of RAM), it can now defer it to a 16 GB backlog because the newer system has 64 GB of RAM.