Workspaces/Advanced Features/Filesystems

Multiple Filesystem Locations

Works on cluster	bwUC 3.0	BinAC2	Helix	JUSTUS 2	NEMO2
-F option (multiple filesystems)	✓	✗	✗	✗	✗

Some clusters offer multiple filesystem locations for workspaces with different characteristics:

bwUniCluster 3.0:

• Default workspace filesystem (Lustre)
• Flash-based workspace filesystem (ffuc) - for KIT/HoreKa users only
  • Lower latency and better performance for small files
  • SSDs instead of hard disks
  • Shared between bwUniCluster 3.0 and HoreKa

Example creating workspace on flash filesystem:

  $ ws_allocate -F ffuc myworkspace 60

Use ws_list -l or ws_find -l to see available filesystem locations on your cluster.

Choosing the Right Filesystem

Note: Most clusters have only one default filesystem - the -F option is rarely needed. Use ws_list -l first to check if multiple filesystems are available on your cluster.

bwUniCluster 3.0 Filesystems

Default Lustre filesystem:

• Standard workspace location
• Best for large files and sequential I/O
• General-purpose storage

Flash filesystem (ffuc):

• SSD-based storage for KIT/HoreKa users
• Shared between bwUniCluster 3.0 and HoreKa
• Use for workloads with:
  • Many small files
  • Random I/O patterns
  • AI/ML training
  • Compilation and builds
• Balance load: use -F ffuc when appropriate to reduce load on default filesystem

General Guidelines

Flash-based filesystems (SSD/NVMe):

• Use for many small files
• Best for low-latency requirements
• Ideal for random I/O patterns
• Examples: Weka (NEMO2), ffuc (bwUniCluster 3.0)

Standard Lustre/parallel filesystems:

• Best for large files
• Optimized for sequential I/O patterns
• General-purpose workload support

Data Type Recommendations

Large sequential I/O:

• Use standard workspace filesystem
• Lustre: best for very large files
• Weka: excellent for both large and small files

Many small files or random access:

• Use flash-based workspace filesystem (Weka, ffuc)
• Or stage to $TMPDIR on compute nodes

Data read multiple times on single node:

• Copy to $TMPDIR at job start for best performance

Temporary data for single node:

• Always use $TMPDIR, not workspaces

Multi-node temporary data:

• Use workspaces (not suitable for $TMPDIR)

AI/ML training data:

• Use Weka (NEMO2) or flash filesystems for best performance
• Or stage to $TMPDIR for repeated access

Compilation/build directories:

• Use flash-based filesystems (Weka, ffuc)
• Or $TMPDIR for better performance

For more information about specific filesystems, see the Quotas & Limits page.