On Thu, Jun 09, 2022 at 03:31:02PM +0100, Richard W.M. Jones wrote:
On Thu, Jun 09, 2022 at 08:34:45AM -0500, Eric Blake wrote:
After the work in the previous patches, it is now a trivial feature addition to support any buffer-like object as the argument to h.aio_{pread,pwrite}, and matching how h.pwrite already did that. For example, you can do h.aio_pwrite(b'123', 0), instead of having to copy into nbd.Buffer first. More importantly, whereas nbd.Buffer.to_bytearray() currently copies out and nbd.Buffer.from_bytearray(buffer) copies in, using native types means you can reduce that copy overhead.
For a demonstration of the speedups possible, compare two scripts that utilize a new buffer for every I/O (important if you are going to do parallel operations - even though this demo is serialized), one using nbd.Buffer, the other using native python types, with timing on my machine:
$ export script1='
$ nbdkit -U - --filter=checkwrite pattern 10G \ --run 'nbdsh -u "$uri" -c "h.set_pread_initialize(True)" -c "script1"'
Should that be "$script1" ? (and same below)
Yep. Serves me right for copy-pasting the wrong line.
Acked-by: Richard W.M. Jones <rjones@redhat.com>
Thanks for looking at this series; I can see from the difference between Acked-by and Reviewed-by that it gets pretty technical ;) -- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3266 Virtualization: qemu.org | libvirt.org