Which directory should I use to analyze many small files?
An interactive batch job on Puhti allows you to start a shell session on a compute node with up to:
- 8 cores
- 76 GB of memory
- 7 days of run time
- 720 GB of fast local scratch disk
To launch an interactive session on Puhti, execute the command:
One of the useful features of interactive batch jobs on Puhti is the NVMe-based
fast local scratch area
($LOCAL_SCRATCH
) that you may request. The "normal" Lustre-based
project-specific directories scratch and projappl can store large amounts
of data and are accessible to all nodes of Puhti. However, these directories
are
inefficient for managing thousands of files.
Generally, you should avoid workflows that require creating large amounts of small files. If you anyhow need to work with thousands of files, you should consider using the NVMe-based temporary local disk areas, either through normal or interactive batch jobs. The local scratch area is node-specific and visible only to the batch job that requests it. After the job ends, the local scratch is purged. Because of this, you need to always first import your dataset to the local disk, and when you finish, copy the data you want to preserve back to some more permanent storage place such as scratch or Allas.
To demonstrate the efficiency of the local scratch area, let's study a sample
directory called big_data
. The directory contains about 100 GB of data in
120 000 files. Initially, the data is packaged as a single tar archive in the
scratch directory of project 2001234 (/scratch/project_2001234/big_data.tar
)
First, we launch an interactive batch job with 2 cores, 4000 MB of memory and 250 GB of fast local temporary disk space.
The analysis is then done in three steps:
-
Move to the local scratch area using the environment variable
$LOCAL_SCRATCH
and extract the tar package: -
Run the analysis. This time we run a
In this example, about 52 000 fasta files were found. Thus, after the processing, thefor
loop that uses a commandtranseq
to translate all fasta files found in thebig_data
directory into new protein sequence files:big_data
directory now contains 52 000 more small files. The actual translation is a computationally light task, so most time is used just to open and close files (I/O). -
When the processing is finished, we store the results back to the scratch directory as a new tar file:
Now the results are safe in one file on scratch, and we can exit the interactive session.
The same analysis procedure could be done in the scratch directory too.
However, it is slow and may degrade the performance of the shared file system
for all users. To demonstrate the difference, below is an execution time
comparison for running the three steps above in $LOCAL_SCRATCH
and in a
regular Lustre-based scratch directory. The response times of $LOCAL_SCRATCH
are rather stable, but in the scratch directory the execution times will vary
a lot based on the current total load on the Lustre file system.
Step | $LOCAL_SCRATCH |
/scratch |
---|---|---|
1. Opening tar file | 2m 8s | 4m 12s |
2. Processing | 9m 42s | 21m 58s |
3. Creating new tar file | 2m 25s | 42m 21s |
Total | 14m 15s | 1h 8m 31s |