[Libguestfs] [PATCH nbdkit] Experiment with parallel python plugin
Nir Soffer
nsoffer at redhat.com
Thu Aug 6 22:06:11 UTC 2020
On Thu, Aug 6, 2020 at 11:52 PM Richard W.M. Jones <rjones at redhat.com> wrote:
>
> On Thu, Aug 06, 2020 at 11:22:00PM +0300, Nir Soffer wrote:
> > This is a quick hack to experiment with parallel threading model in the
> > python plugin.
> >
> > Changes:
> >
> > - Use aligned buffers to make it possible to use O_DIRECT. Using
> > parallel I/O does not buy us much when using buffered I/O. pwrite()
> > copies data to the page cache, and pread() reads data from the page
> > cache.
>
> O_DIRECT is unfortunately a bit too fragile to consider using
> routinely. But I wonder if one of the posix_fadvise(2) hints could be
> used (eg. POSIX_FADV_SEQUENTIAL + POSIX_FADV_DONTNEED). Adding
> fadvise hints as a parameter for the file plugin is a very plausible
> change.
This does not help much, and does not replace direct I/O when using
shared storage.
In oVirt we alway use direct I/O for image copies, for several reasons:
- Reading using buffered I/O you may get stale data form page cache, after image
was modified on another host
- Reading and writing pollute the page cache with data we will never
need, since VM
also use direct I/O.
- Writing using buffered I/O lead to I/O delays in other programs when
the kernel flushes
gigabytes from the page cache to storage. This easily cause timeouts
in sanlock leases
renewal (using 10 seconds timeout), and these timeouts can cause
leases to expire, killing
VMs holding the expired leases, or in the worst case rebooting the
host if sanlock could
not kill a program using an expired lease.
- Performance using direct I/O is usually better and always more consistent
For nbdkit this is nice to have. If we change the buffer allocation to
create aligned
buffers, plugins like the file plugin can add direct=false option to
use direct I/O.
Supporting unaligned requests will not be easy, but a plugin can
reject such requests.
> > - Disable extents in the file plugin. This way we can compare it with
> > the python file example.
> >
> > - Implement flush in the file example.
> >
> > With these changes, I could compare the file plugin with the new python
> > file example, and it seems that the parallel threading models works
> > nicely, and we get similar performance for the case of fully allocated
> > image.
> >
> > I created a test image using:
> >
> > $ virt-builder fedora-32 -o /var/tmp/fedora-32.raw --root-password=password:root
> >
> > And a fully allocated test image using:
> >
> > $ fallocate --length 6g /var/tmp/disk.raw
> > $ dd if=/var/tmp/fedora-32.raw bs=8M of=/var/tmp/disk.raw iflag=direct oflag=direct conv=fsync,notrunc
> >
> > $ qemu-img map --output json /var/tmp/disk.raw
> > [{ "start": 0, "length": 6442450944, "depth": 0, "zero": false, "data": true, "offset": 0}]
> >
> > For reference, copying this image with dd using direct I/O:
> >
> > $ dd if=/var/tmp/disk.raw bs=2M of=/dev/shm/disk.raw iflag=direct conv=fsync status=progress
> > 6442450944 bytes (6.4 GB, 6.0 GiB) copied, 10.4783 s, 615 MB/s
> >
> > Copying same image with qemu-img convert, disabling zero detection,
> > using different number of coroutines:
> >
> > $ time qemu-img convert -f raw -O raw -T none -S0 -m1 -W /var/tmp/disk.raw /dev/shm/disk.raw
> >
> > real 0m11.527s
> > user 0m0.102s
> > sys 0m2.330s
> >
> > $ time qemu-img convert -f raw -O raw -T none -S0 -m2 -W /var/tmp/disk.raw /dev/shm/disk.raw
> >
> > real 0m5.971s
> > user 0m0.080s
> > sys 0m2.749s
> >
> > $ time qemu-img convert -f raw -O raw -T none -S0 -m4 -W /var/tmp/disk.raw /dev/shm/disk.raw
> >
> > real 0m3.674s
> > user 0m0.071s
> > sys 0m3.140s
> >
> > $ time qemu-img convert -f raw -O raw -T none -S0 -m8 -W /var/tmp/disk.raw /dev/shm/disk.raw
> >
> > real 0m3.408s
> > user 0m0.069s
> > sys 0m3.813s
> >
> > $ time qemu-img convert -f raw -O raw -T none -S0 -m16 -W /var/tmp/disk.raw /dev/shm/disk.raw
> >
> > real 0m3.305s
> > user 0m0.054s
> > sys 0m3.767s
> >
> > Same with the modified file plugin, using direct I/O and without
> > extents:
> >
> > $ rm -f /tmp/nbd.sock && ./nbdkit -U /tmp/nbd.sock -t1 -f -r file file=/var/tmp/disk.raw
> > $ time qemu-img convert -f raw -O raw -S0 -m16 -W nbd:unix:/tmp/nbd.sock /dev/shm/disk.raw
> >
> > real 0m12.167s
> > user 0m5.798s
> > sys 0m2.477s
> >
> > $ rm -f /tmp/nbd.sock && ./nbdkit -U /tmp/nbd.sock -t2 -f -r file file=/var/tmp/disk.raw
> > $ time qemu-img convert -f raw -O raw -S0 -m16 -W nbd:unix:/tmp/nbd.sock /dev/shm/disk.raw
> >
> > real 0m7.981s
> > user 0m5.204s
> > sys 0m2.740s
> >
> > $ rm -f /tmp/nbd.sock && ./nbdkit -U /tmp/nbd.sock -t4 -f -r file file=/var/tmp/disk.raw
> > $ time qemu-img convert -f raw -O raw -S0 -m16 -W nbd:unix:/tmp/nbd.sock /dev/shm/disk.raw
> >
> > real 0m6.568s
> > user 0m4.996s
> > sys 0m3.167s
> >
> > $ rm -f /tmp/nbd.sock && ./nbdkit -U /tmp/nbd.sock -t8 -f -r file file=/var/tmp/disk.raw
> > $ time qemu-img convert -f raw -O raw -S0 -m16 -W nbd:unix:/tmp/nbd.sock /dev/shm/disk.raw
> >
> > real 0m6.493s
> > user 0m4.950s
> > sys 0m3.492s
> >
> > $ rm -f /tmp/nbd.sock && ./nbdkit -U /tmp/nbd.sock -t16 -f -r file file=/var/tmp/disk.raw
> > $ time qemu-img convert -f raw -O raw -S0 -m16 -W nbd:unix:/tmp/nbd.sock /dev/shm/disk.raw
> >
> > real 0m6.138s
> > user 0m4.621s
> > sys 0m3.550s
> >
> > Finally, same with the python file example:
> >
> > $ rm -f /tmp/nbd.sock && ./nbdkit -U /tmp/nbd.sock -t1 -f -r python ./plugins/python/examples/file.py file=/var/tmp/disk.raw
> > $ time qemu-img convert -f raw -O raw -S0 -m16 -W nbd:unix:/tmp/nbd.sock /dev/shm/disk.raw
> >
> > real 0m12.398s
> > user 0m6.652s
> > sys 0m2.484s
> >
> > $ rm -f /tmp/nbd.sock && ./nbdkit -U /tmp/nbd.sock -t2 -f -r python ./plugins/python/examples/file.py file=/var/tmp/disk.raw
> > $ time qemu-img convert -p -f raw -O raw -S0 -m16 -W nbd:unix:/tmp/nbd.sock /dev/shm/disk.raw
> >
> > real 0m8.169s
> > user 0m5.418s
> > sys 0m2.736s
> >
> > $ rm -f /tmp/nbd.sock && ./nbdkit -U /tmp/nbd.sock -t4 -f -r python ./plugins/python/examples/file.py file=/var/tmp/disk.raw
> > $ time qemu-img convert -p -f raw -O raw -S0 -m16 -W nbd:unix:/tmp/nbd.sock /dev/shm/disk.raw
> >
> > real 0m6.419s
> > user 0m4.891s
> > sys 0m3.103s
> >
> > $ rm -f /tmp/nbd.sock && ./nbdkit -U /tmp/nbd.sock -t8 -f -r python ./plugins/python/examples/file.py file=/var/tmp/disk.raw
> > $ time qemu-img convert -p -f raw -O raw -S0 -m16 -W nbd:unix:/tmp/nbd.sock /dev/shm/disk.raw
> >
> > real 0m6.610s
> > user 0m5.115s
> > sys 0m3.377s
> >
> > $ rm -f /tmp/nbd.sock && ./nbdkit -U /tmp/nbd.sock -t16 -f -r python ./plugins/python/examples/file.py file=/var/tmp/disk.raw
> > $ time qemu-img convert -p -f raw -O raw -S0 -m16 -W nbd:unix:/tmp/nbd.sock /dev/shm/disk.raw
> >
> > real 0m6.093s
> > user 0m4.520s
> > sys 0m3.567s
>
> All pretty excellent for an interpreted programming language.
This test is the best case for the python plugin since without
extents, all reads are 2 MiB.
If we enable extents we will have many small reads, and the file
plugin will be faster.
> > I think this show that the parallel threading model works for the python
> > plugin as good as for the file plugin.
> > ---
> > plugins/file/file.c | 4 ++--
> > plugins/python/examples/file.py | 5 ++++-
> > server/plugins.c | 20 ++++++++++++++------
> > server/threadlocal.c | 7 +++++--
> > 4 files changed, 25 insertions(+), 11 deletions(-)
> >
> > diff --git a/plugins/file/file.c b/plugins/file/file.c
> > index dc99f992..27316b9f 100644
> > --- a/plugins/file/file.c
> > +++ b/plugins/file/file.c
> > @@ -170,7 +170,7 @@ file_open (int readonly)
> > return NULL;
> > }
> >
> > - flags = O_CLOEXEC|O_NOCTTY;
> > + flags = O_CLOEXEC|O_NOCTTY|O_DIRECT;
> > if (readonly)
> > flags |= O_RDONLY;
> > else
> > @@ -551,7 +551,7 @@ file_can_extents (void *handle)
> > nbdkit_debug ("extents disabled: lseek: SEEK_HOLE: %m");
> > return 0;
> > }
> > - return 1;
> > + return 0;
> > }
> >
> > static int
> > diff --git a/plugins/python/examples/file.py b/plugins/python/examples/file.py
> > index 866b8244..3652eb52 100644
> > --- a/plugins/python/examples/file.py
> > +++ b/plugins/python/examples/file.py
> > @@ -49,7 +49,7 @@ def open(readonly):
> > flags = os.O_RDONLY
> > else:
> > flags = os.O_RDWR
> > - fd = os.open(filename, flags)
> > + fd = os.open(filename, flags | os.O_DIRECT)
> > return { 'fd': fd }
> >
> > def get_size(h):
> > @@ -65,3 +65,6 @@ def pwrite(h, buf, offset, flags):
> > n = os.pwritev(h['fd'], [buf], offset)
> > if n != len(buf):
> > raise RuntimeError("short write")
> > +
> > +def flush(h, flags):
> > + os.fsync(h['fd'])
> > diff --git a/server/plugins.c b/server/plugins.c
> > index d4364cd2..ce4700a3 100644
> > --- a/server/plugins.c
> > +++ b/server/plugins.c
> > @@ -631,6 +631,8 @@ plugin_zero (struct backend *b, void *handle,
> > bool fast_zero = flags & NBDKIT_FLAG_FAST_ZERO;
> > bool emulate = false;
> > bool need_flush = false;
> > + void *zero_buffer = NULL;
> > + int buffer_size = MIN (MAX_REQUEST_SIZE, count);
> >
> > if (fua && backend_can_fua (b) != NBDKIT_FUA_NATIVE) {
> > flags &= ~NBDKIT_FLAG_FUA;
> > @@ -669,19 +671,25 @@ plugin_zero (struct backend *b, void *handle,
> > threadlocal_set_error (0);
> > *err = 0;
> >
> > + *err = posix_memalign(&zero_buffer, 4096, buffer_size);
> > + if (*err != 0) {
> > + r = -1;
> > + goto done;
> > + }
> > +
> > + memset(zero_buffer, 0, buffer_size);
> > +
> > while (count) {
> > - /* Always contains zeroes, but we can't use const or else gcc 9
> > - * will use .rodata instead of .bss and inflate the binary size.
> > - */
> > - static /* const */ char buf[MAX_REQUEST_SIZE];
> > - uint32_t limit = MIN (count, sizeof buf);
> > + uint32_t limit = MIN (count, buffer_size);
> >
> > - r = plugin_pwrite (b, handle, buf, limit, offset, flags, err);
> > + r = plugin_pwrite (b, handle, zero_buffer, limit, offset, flags, err);
> > if (r == -1)
> > break;
> > count -= limit;
> > }
> >
> > + free(zero_buffer);
> > +
> > done:
> > if (r != -1 && need_flush)
> > r = plugin_flush (b, handle, 0, err);
> > diff --git a/server/threadlocal.c b/server/threadlocal.c
> > index 90230028..04c82842 100644
> > --- a/server/threadlocal.c
> > +++ b/server/threadlocal.c
> > @@ -195,13 +195,16 @@ threadlocal_buffer (size_t size)
> >
> > if (threadlocal->buffer_size < size) {
> > void *ptr;
> > + int err;
> >
> > - ptr = realloc (threadlocal->buffer, size);
> > - if (ptr == NULL) {
> > + err = posix_memalign (&ptr, 4096, size);
> > + if (err != 0) {
> > nbdkit_error ("threadlocal_buffer: realloc: %m");
> > return NULL;
> > }
> > +
> > memset (ptr, 0, size);
> > + free(threadlocal->buffer);
> > threadlocal->buffer = ptr;
> > threadlocal->buffer_size = size;
> > }
> > --
> > 2.25.4
>
> Rich.
>
> --
> Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones
> Read my programming and virtualization blog: http://rwmj.wordpress.com
> virt-builder quickly builds VMs from scratch
> http://libguestfs.org/virt-builder.1.html
>
More information about the Libguestfs
mailing list