Instructor Notes

Timing

Episode Teaching Exercises
Why copying hurts 10 0
Pack, then checksum 10 15
rsync 10 15
Several streams (optional) 15 10
From outside the cluster 10 5
Stage in, stage out 10 20

Several streams is optional. Skip it when the room is tight, and keep the pack, rsync, and stage exercises.

Deck

Branch feat/data-movement-deck, commit f7ed5319a86448ac8dda3cb60fe4f58f64b5ca09.

The file is slides-data-movement/data-movement.org, Rohit Goswami, 26 October 2026. PROVENANCE.md records the same pair. The three scripts under episodes/files/ are the sequences on the slides.

Prep

Give the room an empty directory on scratch before predict.sh. The script creates 20,000 one-line files plus a tar archive. Home is the wrong disk for that.

The ext4 note on the deck (26 October 2026) says the order matches the prediction, the partial file was 4128768 bytes of 50331648, and sha256sum -c printed src.tar: OK. It does not record the three real times. Ask the room to write down their own real line. Do not fill in seconds.

stage.slurm is a ten-minute job. Submit it from scratch. The copy back is the second rsync, and it has to run inside that script.

The published scripts name the rome partition, and stage.slurm also sets --constraint=scratch-node. Leave those lines as published. The site tabs are what you show for the machine in the welcome note. The course cluster is a SURF Research Cloud workspace, and the course tab requests --partition=rome.

EESSI 2023.06 and 2025.06 do not provide mpiFileUtils. The course cluster has it installed. That install is not an EESSI module. The course tab shows the commands.

fpsync needs fpart. If it is absent, teach the xargs -P 4 form and move on.

mksquashfs may be absent. The timed exercise uses tar.

Common mistakes

  • Treating a slow cp -r of the 20,000 files as a network problem. The metadata server is doing a create for every file.
  • Running predict.sh in the lesson directory or in home.
  • Dropping the trailing slash on the rsync source and then looking for files one directory too deep.
  • Reading part/big.bin after timeout as a finished copy. cmp is the check, and it prints nothing when the bytes match.
  • Putting the copy back in a second job. The node disk is already gone.
  • Opening four streams on the tiny files. Pack those first. Four streams are for large shards.
  • Expecting scp to continue a partial file. It starts again from the first file.
  • Assuming /archive is visible on a compute node. daget runs on a login or staging node. The deck’s copy onto scratch uses the staging partition.

Why copying hurts

Pack, then checksum

rsync

Several streams

Instructor Note

This episode is optional.


From outside the cluster

Stage in, stage out