[pbs-devel] [PATCH proxmox-backup 2/2] tape/helpers/snapshot_reader: sort chunks by inode (per index)

[Thomas Lamprecht]

On 18.06.21 11:29, Dominik Csapak wrote:
> sort the chunks we want to backup to tape by inode, to gain some
> speed on spinning disks. this is done per index, not globally.
> 
> costs a bit memory, but not too much, about 16 bytes per chunk which
> would mean ~4MiB for a 1TiB index with 4MiB chunks.
> 
> Signed-off-by: Dominik Csapak <d.csapak at proxmox.com>
> ---
> this resulted in a speedup in my setup of between 20 and 30%
> (single spinner with random snapshots, from 17-26MiB/s to 30-40MiB/s)
> we already do this for verification, but got no real feedback on it,
> so either it does not make that much of a difference in the real world,
> or it is not that visible on verification (since that varies very much
> anyway)

or, IMO more likely, people just don't give feedback if things are working out
OK-ish ;-)

> 
>  src/tape/helpers/snapshot_reader.rs | 20 +++++++++++++-------
>  1 file changed, 13 insertions(+), 7 deletions(-)
> 
> diff --git a/src/tape/helpers/snapshot_reader.rs b/src/tape/helpers/snapshot_reader.rs
> index 7b272e37..416c88c1 100644
> --- a/src/tape/helpers/snapshot_reader.rs
> +++ b/src/tape/helpers/snapshot_reader.rs
> @@ -107,7 +107,7 @@ impl SnapshotReader {
>  pub struct SnapshotChunkIterator<'a> {
>      snapshot_reader: &'a SnapshotReader,
>      todo_list: Vec<String>,
> -    current_index: Option<(Arc<Box<dyn IndexFile>>, usize)>,
> +    current_index: Option<(Arc<Box<dyn IndexFile + Send>>, usize, Vec<(usize, u64)>)>,
>  }
>  
>  impl <'a> Iterator for SnapshotChunkIterator<'a> {
> @@ -119,20 +119,26 @@ impl <'a> Iterator for SnapshotChunkIterator<'a> {
>                  if self.current_index.is_none() {
>                      if let Some(filename) = self.todo_list.pop() {
>                          let file = self.snapshot_reader.open_file(&filename)?;
> -                        let index: Box<dyn IndexFile> = match archive_type(&filename)? {
> +                        let index: Box<dyn IndexFile + Send> = match archive_type(&filename)? {
>                              ArchiveType::FixedIndex => Box::new(FixedIndexReader::new(file)?),
>                              ArchiveType::DynamicIndex => Box::new(DynamicIndexReader::new(file)?),
>                              _ => bail!("SnapshotChunkIterator: got unknown file type - internal error"),
>                          };
> -                        self.current_index = Some((Arc::new(index), 0));
> +
> +                        let datastore =
> +                            DataStore::lookup_datastore(self.snapshot_reader.datastore_name())?;
> +                        let order = datastore.get_chunks_in_order(&index, |_| false, |_| Ok(()))?;
> +
> +                        self.current_index = Some((Arc::new(index), 0, order));
>                      } else {
>                          return Ok(None);
>                      }
>                  }
> -                let (index, pos) = self.current_index.take().unwrap();
> -                if pos < index.index_count() {
> -                    let digest = *index.index_digest(pos).unwrap();
> -                    self.current_index = Some((index, pos + 1));
> +                let (index, pos, list) = self.current_index.take().unwrap();
> +                if pos < list.len() {
> +                    let (real_pos, _) = list[pos];
> +                    let digest = *index.index_digest(real_pos).unwrap();
> +                    self.current_index = Some((index, pos + 1, list));
>                      return Ok(Some(digest));
>                  } else {
>                      // pop next index
>