hole punching performance

[Dave Chinner]

On Wed, Jan 02, 2013 at 04:51:07PM -0500, Bradley C. Kuszmaul wrote:
> If I use hole-punching, what will happen to the performance of my application?
> 
> I have a multithreaded application that creates large files (many
> gigabytes per file).  The application sometimes wants to punch holes
> (say 1 megabyte in size).
> 
> On Redhat 6, I've measured that punching holes requires about 2ms

What version of RHEL 6.x? On x <= 1, hole punching is a synchronous
transaction.  On x >= 2, it is an asynchronous transaction and so is
much, much faster.

> (this with a battery-backed up RAID controller), which is slower than
> I was hoping for, but it's probably OK.  The throughput is only about
> 2ms per hole-punch even if I have lots of threads punching holes in
> lots of different files at the same time.

That sounds like synchronous transaction behaviour.

A current 3.8-rc1 kernel does a hole punch in well under 2ms. Here's
10,000 hole punches being done in ~300ms:

$ cat t.c
#define _GNU_SOURCE
#include <unistd.h>
#include <fcntl.h>
#include <linux/falloc.h>
#include <xfs/xfs.h>

int main(int argc, char *argv[])
{
        int i, fd;
        fd = open("/mnt/scratch/blah", O_CREAT|O_TRUNC|O_RDWR, 0777);
        perror("open");
        fallocate(fd, 0, 0, 20 * 1024 * 1024 * 1024LL);
        for (i = 0; i < 10000; i++) {
                // fallocate(fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE, i * 8192, 4096);
                struct xfs_flock64      l = {0};

                l.l_whence = SEEK_SET;
                l.l_start = i * 8192;
                l.l_len = 4096;

                ioctl(fd, XFS_IOC_UNRESVSP, &l);
        }
        close(fd);
}
dave at test-4:~$ gcc -O2 t.c
dave at test-4:~$ rm -f /mnt/scratch/blah
dave at test-4:~$ time ./a.out
open: Success

real    0m0.336s
user    0m0.000s
sys     0m0.336s
dave at test-4:~$

So that means roughly 300ms/10000 = 30uS per hole punch call.
I get the same result with fallocate or XFS_IOC_UNRESVSP, and I get
the same result on RHEL 6.2+.

> The question I have:  What will happen to the performance of other
> threads doing read() and write() operations?  Will hole-punching slow
> down the other read() and write() operations running in other threads?

That all depends. Hole punching is serialised the same way as
truncation - all concurrent operations to the same file are locked
out while the hole punch is performed. Operations to other files
will unaffected unless they are trying to allocate or free extents
in the same allocation group, or you are running a kernel that does
synchronous transactions and the other operations serialise on the
synchronous transaction commits...

Cheers,

Dave.
-- 
Dave Chinner
david at fromorbit.com