Hi All.
I saw an many years old thread on this topic and recently ran into the
same problem. I'm wondering if anything has changed since - or someone
had some smart thoughts on this?
In short, if you do:
fd = open("foo",O_DIRECT|O_TRUNC_O|CREAT,0666);
fallocate(fd,0,xxx);
io_submit(.. IO_CMD_PWRITE ..);
io_submit blocks on ext4. (It also blocks on ext3, but fallocate fails
there so that is not entirely surprising..)
On xfs io_submit runs asynchronously with the same sequence. There is
a change in the performance characteristic: -30% with a 32k iosize
compared to re-writing already existing blocks - but it performs well
with larger iosizes
I am working on a network file copy application, so this is the common
(well. only) workload.
I haven't yet got access to the good storage to test on, but I'm
concerned that io_submit blocking means that there is going to be
limited, or no, concurrency at the SCSI level? I understand the
problem is that ext4 needs to note the block is now actually used and
no longer 'zero' but I guess I don't entirely follow why that wrecks
AIO?
Thanks,
Jason