Research Computing Facilities Technical Summary

Boston University Shared Computing Cluster (SCC)

Listed below are the technical details on the SCC login and compute nodes, including run time limits, SU charge rate for each node, and the configuration of the batch system. If for some reason your code is not able to run within these parameters, please don’t hesitate to send email to help@scc.bu.edu.

Hardware Configuration

¹These machines have limited access; not all SCF users can fully utilize these systems. For those users with special access to these systems, the SU charge is 0.0 for these systems only.
²Each of these nodes has two NVIDIA Tesla K40m GPU Cards with 12 GB of Memory.
³Each of these nodes has four NVIDIA Tesla K40m GPU Cards with 12 GB of Memory.
⁴Each of these nodes has two or four (specified above) NVIDIA Tesla P100 GPU Cards with 12 GB of Memory.
⁵This node is for support of VirtualGL and has one NVIDIA Tesla K2200 GPU Card with 4 GB of Memory.
⁶These nodes are for support of VirtualGL and each has one NVIDIA Tesla M2000 GPU Card with 4 GB of Memory.
⁷Each of these nodes has one, two, or four (specified above) NVIDIA Tesla V100 GPU Cards with 16 GB of Memory.
⁸Each of these nodes has one, two, or four (specified above) NVIDIA Tesla V100 GPU Cards with 32 GB of Memory.
⁹Each of these nodes has one, two, or four (specified above) NVIDIA Tesla P100 GPU Cards with 16 GB of Memory.
¹⁰Each of these nodes has the indicated number of NVIDIA Tesla A100 GPU Card with 40 GB of Memory.
¹¹Each of these nodes has the indicated number of NVIDIA A6000 GPU Cards with 48 GB of Memory each.
¹²Each of these nodes has the indicated number of NVIDIA RTX 8000 GPU Cards with 48 GB of Memory each.
¹³Each of these nodes has the indicated number of NVIDIA TitanXp GPU Cards with 12 GB of Memory each.
¹⁴Each of these nodes has the indicated number of NVIDIA TitanV GPU Cards with 12 GB of Memory each.
¹⁵Each of these nodes has the indicated number of NVIDIA RTX 6000 GPU Cards with 24 GB of Memory each.
¹⁶Each of these nodes has the indicated number of NVIDIA A40 GPU Cards with 48 GB of Memory each.
¹⁷Each of these nodes has the indicated number of NVIDIA Tesla A100 GPU Cards with 80 GB of Memory.
¹⁸Each of these nodes has the indicated number of NVIDIA L40 GPU Cards with 48 GB of Memory each.
¹⁹Each of these nodes has the indicated number of NVIDIA RTX 6000 ada GPU Cards with 48 GB of Memory each.
²⁰Each of these nodes has the indicated number of NVIDIA L40S GPU Cards with 48 GB of Memory each.

Use of all of the GPUs is currently in ‘friendly user’ mode so only the main CPU usage is charged at the moment for all SCC nodes with GPUs.

The Knights Landing nodes (scc-ib1 and scc-ib2) are also currently in ‘friendly user’ mode with no charge for CPU usage on these nodes.

Depending on the speed, memory, and other factors, each node type is charged at a different SU rate per hour, as shown in the table above. Our allocations policy is explained in more detail here.

Batch System and Usage

The batch system on the SCC is the Open Grid Scheduler (OGS), which is an open source batch system based on the Sun Grid Engine scheduler.

Job Run Time Limits
The limitations below apply to the shared SCC resources. Limitations on buy-In nodes are defined by their owners.

Limit	Description
15 minutes of CPU Time	Jobs running on the login nodes are limited to 15 minutes of CPU time and a small number of cores.
12 hours default wall clock	Jobs on the batch nodes have a default wall clock limit of 12 hours but this can be increased, depending on the type of job. Use the qsub option -l h_rt=HH:MM:SS to ask for a higher limit.
720 hours -serial job 720 hours – omp (1 node) job 120 hours – mpi job 48 hours – GPU job	Single processor (serial) and omp (multiple cores all on one node) jobs can run for 720 hours, MPI jobs for 120 hours, and jobs using GPUs are limited to 48 hours.
1024 cores	An individual user is also only allowed to have 1024 shared cores maximum simultaneously in the run state. This limit does not affect job submission.

Note that usage on the SCC is charged by wall clock time; if you request 8 cores for 12 hours and your code runs for 3 hours, you will be charged for the 24 hours (8 cores * 3 hours) even if your job did not use all requested cores. The charge is computed by multiplying the wall clock time by the SU factor of the node(s) used.

---