Ftruncate vs fallocate

(Clement Pang) #1

We are noticing (starting in 5.2.x) that fdb trace logs are showing Net2SlowTaskTrace truncate() in which points to

AsyncFileKAIO::truncate(long) at /home/ubuntu/foundationdb/fdbrpc/
AsyncFileKAIO::launch() at /home/ubuntu/foundationdb/./flow/flow.h:636
 (inlined by) AsyncFileKAIO::launch() at /home/ubuntu/foundationdb/fdbrpc/
N2::Net2::run() at /home/ubuntu/foundationdb/flow/
main at /home/ubuntu/foundationdb/fdbserver/ (discriminator 2)

The specific line that’s slow is fallocate(). In 3.x, the code does a simple ftruncate(fd, size) while in 5.2.x it attempts a fallocate() if it’s supported (which it is on our system, the code would shut that off anyways and do ftruncate() anyways). Just throwing it out there to see if that’s something that’s known to happen (and if it’s an issue with the file system being inefficient with fallocate() vs ftruncate()). It does show up as an increase in CPU usage though. Will probably attempt a custom build to test the differences between the two.

(A.J. Beamon) #2

I do know I’ve seen similar slow tasks, although I was thinking they were also present when it was formerly a call to ftruncate. I’m not really sure about that, and I don’t think we’ve done much investigation into these slow tasks, so if you find anything interesting please share it with us.

(Clement Pang) #3

Will do. We are running with ftruncate now, from what I can read, fallocate should be faster but i guess it depends on the implementation.