On Fri, Aug 7, 2020, 16:16 Richard W.M. Jones <rjones@redhat.com> wrote:
On Fri, Aug 07, 2020 at 07:53:13AM -0500, Eric Blake wrote:
> >$ free -m; time ./nbdkit file /var/tmp/random fadvise=sequential cache=none --run 'qemu-img convert -n -p -m 16 -W $nbd "json:{\"file.driver\":\"null-co\",\"file.size\":\"1E\"}"' ; free -m ; cachestats /var/tmp/random
>
> Hmm - the -W actually says that qemu-img is performing semi-random
> access (there is no guarantee that the 16 coroutines are serviced in
> linear order of the file), even though we really are making only one
> pass through the file in bulk.  I don't know if fadvise=normal would
> be any better; dropping -W but keeping -m 16 might also be an
> interesting number to check (where qemu-img tries harder to do
> in-order access, but still take advantage of parallel threads).
>
> >               total        used        free      shared  buff/cache   available
> >Mem:          32083        1188       27928           1        2966       30440
> >Swap:         16135          16       16119
> >     (100.00/100%)
> >
> >real 0m13.107s
> >user 0m2.051s
> >sys  0m37.556s
> >               total        used        free      shared  buff/cache   available
> >Mem:          32083        1196       27861           1        3024       30429
> >Swap:         16135          16       16119
> >pages in cache: 14533/8388608 (0.2%)  [filesize=33554432.0K, pagesize=4K]

Without -W it's very similar:

$ free -m; time ./nbdkit file /var/tmp/random fadvise=sequential cache=none --run 'qemu-img convert -n -p -m 16 $nbd "json:{\"file.driver\":\"null-co\",\"file.size\":\"1E\"}"' ; free -m ; cachestats /var/tmp/random
              total        used        free      shared  buff/cache   available
Mem:          32083        1184       26113           1        4785       30444
Swap:         16135          16       16119
    (100.00/100%)

real    0m13.308s
user    0m1.961s
sys     0m40.455s
              total        used        free      shared  buff/cache   available
Mem:          32083        1188       26049           1        4845       30438
Swap:         16135          16       16119
pages in cache: 14808/8388608 (0.2%)  [filesize=33554432.0K, pagesize=4K]

With -W and using fadvise=random is also about the same:

$ free -m; time ./nbdkit file /var/tmp/random fadvise=random cache=none --run 'qemu-img convert -n -p -m 16 -W $nbd "json:{\"file.driver\":\"null-co\",\"file.size\":\"1E\"}"' ; free -m ; cachestats /var/tmp/random
              total        used        free      shared  buff/cache   available
Mem:          32083        1187       26109           1        4785       30440
Swap:         16135          16       16119
    (100.00/100%)

real    0m13.030s
user    0m1.986s
sys     0m37.498s
              total        used        free      shared  buff/cache   available
Mem:          32083        1187       26053           1        4842       30440
Swap:         16135          16       16119
pages in cache: 14336/8388608 (0.2%)  [filesize=33554432.0K, pagesize=4K]

I'm going to guess that for this case readahead doesn't have much time
to get ahead of qemu.

> >+=item B<fadvise=normal>
> >+
> >+=item B<fadvise=random>
> >+
> >+=item B<fadvise=sequential>
> >+
> >+This optional flag hints to the kernel that you will access the file
> >+normally, or in a random order, or sequentially.  The exact behaviour
> >+depends on your operating system, but for Linux using C<normal> causes
> >+the kernel to read-ahead, C<sequential> causes the kernel to
> >+read-ahead twice as much as C<normal>, and C<random> turns off
> >+read-ahead.
>
> Is it worth a mention of L<posix_fadvise(3)> here, to let the user
> get some idea of what their operating system supports?

Yes I had this at one point but I seem to have dropped it.  Will
add it back, thanks.

> >+=head2 Reducing evictions from the page cache
> >+
> >+If the file is very large and you known the client will only
> >+read/write the file sequentially one time (eg for making a single copy
> >+or backup) then this will stop other processes from being evicted from
> >+the page cache:
> >+
> >+ nbdkit file disk.img fadvise=sequential cache=none
>
> It's also possible to avoid polluting the page cache by using
> O_DIRECT, but that comes with harder guarantees (aligned access
> through aligned buffers), so we may add it as another mode later on.
> But in the meantime, cache=none is fairly nice while still avoiding
> O_DIRECT.

I'm not sure if or even how we could ever do a robust O_DIRECT

We can let the plugin an filter deal with that. The simplest solution is to drop it on the user and require aligned requests.

Maybe a filter can handle alignment?

implementation, but my idea was that it might be an alternate
implementation of cache=none.  But if we thought we might use O_DIRECT
as a separate mode, then maybe we should rename cache=none.
cache=advise?  cache=dontneed?  I can't think of a good name!

Yes, don't call it none if you use the cache.

How about advise=?

I would keep cache semantics similar to qemu.



> >@@ -355,6 +428,17 @@ file_pwrite (void *handle, const void *buf, uint32_t count, uint64_t offset,
> >  {
> >    struct handle *h = handle;
> >+#if defined (HAVE_POSIX_FADVISE) && defined (POSIX_FADV_DONTNEED)
> >+  uint32_t orig_count = count;
> >+  uint64_t orig_offset = offset;
> >+
> >+  /* If cache=none we want to force pages we have just written to the
> >+   * file to be flushed to disk so we can immediately evict them from
> >+   * the page cache.
> >+   */
> >+  if (cache_mode == cache_none) flags |= NBDKIT_FLAG_FUA;
> >+#endif
> >+
> >    while (count > 0) {
> >      ssize_t r = pwrite (h->fd, buf, count, offset);
> >      if (r == -1) {
> >@@ -369,6 +453,12 @@ file_pwrite (void *handle, const void *buf, uint32_t count, uint64_t offset,
> >    if ((flags & NBDKIT_FLAG_FUA) && file_flush (handle, 0) == -1)
> >      return -1;
> >+#ifdef HAVE_POSIX_FADVISE
> >+  /* On Linux this will evict the pages we just wrote from the page cache. */
> >+  if (cache_mode == cache_none)
> >+    posix_fadvise (h->fd, orig_offset, orig_count, POSIX_FADV_DONTNEED);
> >+#endif
>
> So on Linux, POSIX_FADV_DONTNEED after a write that was not flushed
> doesn't help?  You did point out that the use of FUA for flushing
> slows things down, but that's a fair price to pay to keep the cache
> clean.

On Linux POSIX_FADV_DONTNEED won't flush dirty buffers.  I expect (but
didn't actually measure) that just after a medium sized write the
buffers would all be dirty so the posix_fadvise(DONTNEED) call would
do nothing at all.  The advice online does seem to be that you must
flush before calling this.  (Linus advocates a complex
double-buffering solution so that you can be reading into one buffer
while flushing the other, so you don't have the overhead of waiting
for the flush).

I'm going to do a bit of benchmarking of the write side now.

We already tried this with dd and the results were not good.

Nir


Thanks,

Rich.

> Patch looks good to me.
>
> --
> Eric Blake, Principal Software Engineer
> Red Hat, Inc.           +1-919-301-3226
> Virtualization:  qemu.org | libvirt.org

--
Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones
Read my programming and virtualization blog: http://rwmj.wordpress.com
libguestfs lets you edit virtual machines.  Supports shell scripting,
bindings from many languages.  http://libguestfs.org