Sometimes it is needed to copy large number of files from /scratch to /project or vice versa. Both cp and rsync are convenient but sometimes you need speed to do such task.
dcp or distributed copy is a MPI-based copy tool developed by Lawrance Livermore National Lab (LLNL) as part of their mpifileutils suite. We have installed it on Shaheen. Here is an example jobscript to launch a data moving job with dcp:
#!/bin/bash
#SBATCH --ntasks=4
#SBATCH --time=01:00:00
#SBATCH --hint=nomultithread
module load mpifileutils
time srun -n ${SLURM_NTASKS} dcp --verbose --progress 60 --preserve /path/to/source/directory /path/to/destination/directory
|
The above script launches dcp in parallel on with 4 MPI processes. --progress 60 implies that the progress of the operation will be reported every 60 seconds. --preserve implies that the ACL permissions, group ownership, timestamps and extended attributes will be preserved on the files the destination directory as were in parent/source directory.
The following is an example output:
[2021-01-21T16:01:51] Preserving file attributes. [2021-01-21T16:01:51] Walking /project/k01/shaima0d/Shaheen3/Stenchikov/PinatuboInitialStage [2021-01-21T16:01:51] Walking /project/k01/shaima0d/Shaheen3/Stenchikov/README.txt [2021-01-21T16:01:51] Walking /project/k01/shaima0d/Shaheen3/Stenchikov/build_alex.slurm [2021-01-21T16:01:51] Walking /project/k01/shaima0d/Shaheen3/Stenchikov/build_alex2.slurm [2021-01-21T16:01:51] Walking /project/k01/shaima0d/Shaheen3/Stenchikov/build_own.slurm [2021-01-21T16:01:51] Walking /project/k01/shaima0d/Shaheen3/Stenchikov/kuwait_heavy.slurm [2021-01-21T16:01:51] Walking /project/k01/shaima0d/Shaheen3/Stenchikov/minmax.ncl [2021-01-21T16:01:51] Walking /project/k01/shaima0d/Shaheen3/Stenchikov/nasaballon.slurm [2021-01-21T16:01:51] Walking /project/k01/shaima0d/Shaheen3/Stenchikov/run_wrf_371_kuwait_heavy.sh [2021-01-21T16:01:51] Walking /project/k01/shaima0d/Shaheen3/Stenchikov/run_wrf_371_nasaballoon_light.sh [2021-01-21T16:01:51] Walking /project/k01/shaima0d/Shaheen3/Stenchikov/slurm-17933845.out [2021-01-21T16:01:51] Walking /project/k01/shaima0d/Shaheen3/Stenchikov/slurm-17933846.out [2021-01-21T16:01:51] Walking /project/k01/shaima0d/Shaheen3/Stenchikov/slurm-17933847.out [2021-01-21T16:01:51] Walking /project/k01/shaima0d/Shaheen3/Stenchikov/slurm-17933848.out [2021-01-21T16:01:51] Walking /project/k01/shaima0d/Shaheen3/Stenchikov/slurm-17934624.out [2021-01-21T16:01:51] Walking /project/k01/shaima0d/Shaheen3/Stenchikov/submit_script.sh [2021-01-21T16:01:52] Walked 7844 items in 0.595307 secs (13176.397504 items/sec) ... [2021-01-21T16:01:52] Walked 7844 items in 0.595524 seconds (13171.591813 items/sec) [2021-01-21T16:01:52] Copying to /scratch/shaima0d/Shaheen3/Stenchikov [2021-01-21T16:01:52] Items: 7844 [2021-01-21T16:01:52] Directories: 189 [2021-01-21T16:01:52] Files: 7247 [2021-01-21T16:01:52] Links: 408 [2021-01-21T16:01:52] Data: 531.085 GB (75.042 MB per file) [2021-01-21T16:01:52] Creating directories. [2021-01-21T16:01:52] level=6 min=0 max=1 sum=1 rate=272.853500/sec secs=0.003665 [2021-01-21T16:01:52] level=7 min=0 max=5 sum=19 rate=515.727600/sec secs=0.036841 [2021-01-21T16:01:52] level=8 min=0 max=10 sum=59 rate=541.667256/sec secs=0.108923 [2021-01-21T16:01:52] level=9 min=1 max=6 sum=33 rate=556.079307/sec secs=0.059344 [2021-01-21T16:01:52] level=10 min=0 max=19 sum=60 rate=521.802914/sec secs=0.114986 [2021-01-21T16:01:52] level=11 min=0 max=6 sum=12 rate=542.864132/sec secs=0.022105 [2021-01-21T16:01:52] level=12 min=0 max=2 sum=4 rate=555.776195/sec secs=0.007197 [2021-01-21T16:01:52] level=13 min=0 max=1 sum=1 rate=515.207468/sec secs=0.001941 [2021-01-21T16:01:52] level=14 min=0 max=0 sum=0 rate=0.000000/sec secs=0.000001 [2021-01-21T16:01:52] Created 189 directories in 0.355161 seconds (532.153096 items/sec) [2021-01-21T16:01:52] Creating files. [2021-01-21T16:01:52] level=6 min=0 max=6 sum=15 rate=460.022813 secs=0.032607 [2021-01-21T16:01:52] level=7 min=0 max=7 sum=25 rate=471.742915 secs=0.052995 [2021-01-21T16:02:01] level=8 min=141 max=540 sum=3995 rate=434.750857 secs=9.189171 [2021-01-21T16:02:03] level=9 min=1 max=155 sum=516 rate=452.639110 secs=1.139981 [2021-01-21T16:02:07] level=10 min=4 max=382 sum=1763 rate=435.794907 secs=4.045481 [2021-01-21T16:02:09] level=11 min=0 max=260 sum=1039 rate=449.518504 secs=2.311362 [2021-01-21T16:02:10] level=12 min=9 max=66 sum=249 rate=362.368935 secs=0.687145 [2021-01-21T16:02:10] level=13 min=0 max=38 sum=47 rate=416.477838 secs=0.112851 [2021-01-21T16:02:10] level=14 min=0 max=4 sum=6 rate=392.927444 secs=0.015270 [2021-01-21T16:02:10] Created 7655 items in 17.587330 seconds (435.256520 items/sec) [2021-01-21T16:02:10] Copying data. [2021-01-21T16:03:10] Copied 102.215 GB in 60.072 secs (1.702 GB/s) ... [2021-01-21T16:04:10] Copied 202.647 GB in 120.113 secs (1.687 GB/s) ... [2021-01-21T16:05:10] Copied 302.142 GB in 180.145 secs (1.677 GB/s) ... [2021-01-21T16:06:10] Copied 402.684 GB in 240.246 secs (1.676 GB/s) ... [2021-01-21T16:07:51] Copied 499.097 GB in 341.481 secs (1.462 GB/s) ... [2021-01-21T16:07:51] Copied 531.085 GB in 341.482 secs (1.555 GB/s) done [2021-01-21T16:07:51] Copy data: 531.085 GB (570247967642 bytes) [2021-01-21T16:07:51] Copy rate: 1.555 GB/s (570247967642 bytes in 341.481616 seconds) [2021-01-21T16:07:51] Syncing data to disk. [2021-01-21T16:07:52] Sync completed in 0.716662 seconds. [2021-01-21T16:07:52] Setting ownership, permissions, and timestamps. [2021-01-21T16:08:04] Updated 7844 items in 12.315612 seconds (636.915157 items/sec) [2021-01-21T16:08:04] Syncing directory updates to disk. [2021-01-21T16:08:04] Sync completed in 0.055182 seconds. [2021-01-21T16:08:04] Started: Jan-21-2021,16:01:52 [2021-01-21T16:08:04] Completed: Jan-21-2021,16:08:04 [2021-01-21T16:08:04] Seconds: 372.536 [2021-01-21T16:08:04] Items: 7844 [2021-01-21T16:08:04] Directories: 189 [2021-01-21T16:08:04] Files: 7247 [2021-01-21T16:08:04] Links: 408 [2021-01-21T16:08:04] Data: 531.085 GB (570247967642 bytes) [2021-01-21T16:08:04] Rate: 1.426 GB/s (570247967642 bytes in 372.536 seconds) |
As a benchmark, lets try copying 37760 CSV files each of 6.5kB (a total of 241.085 MB).
The table below compares the baseline time taken by cp command to copy these files from /project to /scratch with that taken by dcp with different number of MPI processes:
(MPI) processes | Time to completion | Speedup | |
|---|---|---|---|
| 1 (serial) | 1139.75 seconds | 1 |
| 4 | 888.966 seconds | 1.282 |
| 16 | 226.064 seconds | 5.042 |
| 32 | 401.479 seconds | 2.838 |
Given large enough number of files, you can see decent gains in using dcp. Use >1 Lustre strip when writing big files. This will increase throughput on each file since dcp does not decompose the files into blocks itself.
Throwing more MPI processes may not always give you the more speedup, as seen in the case of 32 vs 16 MPI processes in the table above. Significantly less work (i.e. files to copy) per MPI process can introduce the MPI overhead of synchronizing and slows down the whole job. Thus there is a sweet spot.
It is possible to have variability in time give the time when your copy job runs and the load on the metadata server of Lustre filesystem.