[Libguestfs] [PATCH libnbd 2/2] copy: Set default request-size to 2**18 (262144 bytes)

Sun Jun 20 22:01:10 UTC 2021

On Sun, Jun 20, 2021 at 11:58 PM Richard W.M. Jones <rjones at redhat.com> wrote:
>
> On Sun, Jun 20, 2021 at 10:21:52PM +0300, Nir Soffer wrote:
> > On Sun, Jun 20, 2021 at 7:46 PM Richard W.M. Jones <rjones at redhat.com> wrote:
> > >
> > > As Nir has often pointed out, our current default request buffer size
> > > (32MB) is too large, resulting in nbdcopy being as much as 2˝ times
> > > slower than it could be.
> > >
> > > The optimum buffer size most likely depends on the hardware, and may
> > > even vary over time as machines get generally larger caches.  To
> > > explore the problem I used this command:
> > >
> > > $ hyperfine -P rs 15 25 'nbdkit -U - sparse-random size=100G seed=1 --run "nbdcopy --request-size=\$((2**{rs})) \$uri \$uri"'
> >
> > This uses the same process for serving both reads and writes, which may
> > be different from real world usage when one process is used for reading
> > and one for writing.
>
> Yes, we do need to check whether nbdkit-sparse-random-plugin is a
> realistic test of copying disk images, which I've not really done.
>
> However as far as I know it's not a problem that only having a single
> nbdkit process would be a problem?  (More below ...)
>
> > > On my 2019-era AMD server with 32GB of RAM and 64MB * 4 of L3 cache,
> > > 2**18 (262144) was the optimum when I tested all sizes between
> > > 2**15 (32K) and 2**25 (32M, the current default).
> > >
> > > Summary
> > >   'nbdkit -U - sparse-random size=100G seed=1 --run "nbdcopy --request-size=\$((2**18)) \$uri \$uri"' ran
> > >     1.03 ą 0.04 times faster than 'nbdkit -U - sparse-random size=100G seed=1 --run "nbdcopy --request-size=\$((2**19)) \$uri \$uri"'
> > >     1.06 ą 0.04 times faster than 'nbdkit -U - sparse-random size=100G seed=1 --run "nbdcopy --request-size=\$((2**17)) \$uri \$uri"'
> > >     1.09 ą 0.03 times faster than 'nbdkit -U - sparse-random size=100G seed=1 --run "nbdcopy --request-size=\$((2**20)) \$uri \$uri"'
> >
> > The difference is very small up to this point
>
> I'm going to guess that we bump into some working set vs cache size wall.
> We need to instrument nbdcopy to print the actual working set, since
> it's hard to calculate it from first principles.  (Actually I'm hoping
> Abhay will look at this.)
>
> ...
>
> > But all these results do not test real work copy. They test copying
> > from memory to memory with zero (practical) latency.
>
> (Continuing from above ...)  The problem with doing a practical test
> is that we'd need to set up, eg., 3 machines, and things like network
> bandwidth would likely mask the performance of nbdcopy since I'm
> pretty sure nbdcopy can copy much faster than any network we would
> hope to set up.
>
> Latency could be more of an issue, since it's plausible that on an NBD
> connection with non-zero latency we'd get better performance by using
> more parallel requests.  At least, that's my intuition based on the
> idea that if we have only 1 request at a time, then we suffer a full
> round trip per request, and if that RTT is non-zero then obviously
> more parallelism that one at a time would be better.
>
> The upshot would be that we could end up optimizing for the Unix
> domain socket (near-zero latency) case and find that performance over
> a real network needs different parameters.
>
> But I do like that nbdkit-sparse-random-plugin makes the tests easy to
> do and highly reproducible.
>
> This is surprising:
>
> > When I tested using real storage on a real server, I got best results
> > using 16 requests and one connection and a request size of 1m.
>
> When I've tested multi-conn, more has almost always been better.
>
> > 4 connections with 4 requests per connection with the same request
> > size seem to be ~10% faster in these conditions.
>
> So that's 10% faster than 16 requests x 1 connection?

Yes.

> Anyhow I agree it's a 3D (or higher dimension) space that needs to be
> explored.
>
> > I posted more info on these tests here:
> > https://listman.redhat.com/archives/libguestfs/2021-May/msg00124.html
>
> IIUC there is one high end server.  Is both the source and destination
> located on the NetApp SAN?  It wasn't clear.

The source was on local fast nvme, so this test was not really complete.
I use this setup since I had it for testing imageio upload and download
which is always done to/from a local file or NFS (I did not have proper
NFS server connected to this setup).

> Also since it's been
> about 15 years since I touched FC, what's the bandwidth/latency like
> on modern systems?

I think the setup I tested had 4 or 8 Gbit HBAs, so it was not a very high
end system.

> I'm concerned that using a SAN could just be a
> bottleneck to testing, and we'd be better off testing everything with
> local NVMe.

I think it is best to test with both shared (nfs, fc, iscsi, ceph?) and local
storage. In the best case optimization for local storage will not harm
performance on shared storage.

> > Of course testing with other servers or storage can show different results,
> > and it is impossible to find a value that will work best in all cases.
> >
> > I think we need to test both the number of requests and connections to
> > improve the defaults.
>
> Agreed.
>
> > > ---
> > >  copy/main.c | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > >
> > > diff --git a/copy/main.c b/copy/main.c
> > > index 0fddfc3..70534b5 100644
> > > --- a/copy/main.c
> > > +++ b/copy/main.c
> > > @@ -50,7 +50,7 @@ bool flush;                     /* --flush flag */
> > >  unsigned max_requests = 64;     /* --requests */
> > >  bool progress;                  /* -p flag */
> > >  int progress_fd = -1;           /* --progress=FD */
> > > -unsigned request_size = MAX_REQUEST_SIZE;  /* --request-size */
> > > +unsigned request_size = 1<<18;  /* --request-size */
> >
> > But this is clearly a better default.
> >
> > >  unsigned sparse_size = 4096;    /* --sparse */
> > >  bool synchronous;               /* --synchronous flag */
> > >  unsigned threads;               /* --threads */
> > > --
> > > 2.32.0
>
> I think I'm going to push this, and then I'll discuss with Abhay if
> it's possible to explore more questions:
>
>  * Comparing and characterising nbdkit-sparse-random-plugin vs
>    "real" copies.
>
>  * Exploring #connections vs #requests vs #request size and seeing if
>    there's a better minimum or set of recommendations.
>
> Rich.
>
> --
> Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones
> Read my programming and virtualization blog: http://rwmj.wordpress.com
> virt-df lists disk usage of guests without needing to install any
> software inside the virtual machine.  Supports Linux and Windows.
> http://people.redhat.com/~rjones/virt-df/
>