Tigressdata

Tigressdata is a single computer with 80 hardware threads (40 cores), 768 GB of RAM and an NVIDIA P100 GPU. The system is intended for data analysis and remote visualization of data generated on the HPC clusters. Tigressdata can also be used for developing, debugging and testing codes.

System Configuration and Usage

General Guidelines

As its name implies, tigressdata has fiber connectivity to /tigress (and /projects), which is the large archival storage system. There is also NFS connectivity to selected parallel or scratch storage spaces allocated to the Princeton clusters. Several commercial and open-source packages are installed on tigressdata.

Please be mindful that tigressdata is a shared resource for all users (i.e., there is no job scheduler). Use the htop command and see "Job Scheduling" below to monitor usage.

Accessing Files on Della, perseus and tiger

The figure below makes it clear that the /scratch/gpfs filesystems of della, perseus and tiger are accessible from tigressdata:

For example, to access files on della's /scratch/gpfs filesystem from tigressdata use the path /della/scratch/gpfs/<YourNetID>:

$ ssh <YourNetID>@tigressdata.princeton.edu
$ ls /della/scratch/gpfs/<YourNetID>

The commands above also apply to perseus and tiger with the appropriate changes.

Hardware Configuration
  Processor Speed Nodes Cores per Node Memory per Node Total Cores Interconnect Performance: Theoretical
Tigressdata
Dell Linux Server
2.4 GHz Gold 6148 Xeon 1 20 768 GB 40 N/A 400+ GFLOPS
Job Scheduling

There is no batch scheduler running on tigressdata, and while the only usage limit imposed is a 250 GB memory limit per process, users are urged to avoid launching processes that will overburden the system. Before launching any memory- or compute-intensive tasks, please check

cat /proc/loadavg
cat /proc/meminfo
The first three values in the loadavg output are running averages of the system load for the last one, five and 15 minutes. Oversubscribing the system with CPU-bound tasks will severely compromise throughput for all users. The MemFree field of the meminfo output is the most important to check before launching a task with a large memory footprint. However, if the SwapFree is significantly less than the SwapTotal value, then performance is already compromised, and adding another memory-intensive task will only exacerbate the problem.