ARCHIVED: On the TeraGrid, how are compute jobs charged?
Note: As part of the transition to the Extreme Science and Engineering Discovery Environment (XSEDE):
- Pople (PSC), will retire from TeraGrid service August 31,
2011. Pople users can log in until September 16, 2011, to move files
to the PSC
File Archiver (
golem.psc.edu). Also, Pople users now have accounts on Blacklight (PSC).
- The Condor Pool (Purdue) and Steele (Purdue) will be available to
TeraGrid users through June 30, 2013.
- Several TeraGrid systems will continue to operate as part of XSEDE. Check with system administrators to determine if you should apply for an XSEDE allocation.
Computational resources on the TeraGrid are allocated and charged in service units (SUs). SUs are defined locally on each system, with conversion factors among systems based on HPL benchmark results (see the TeraGrid SU Conversion Calculator).
Current TeraGrid supercomputers have complex multi-core and memory hierarchies. Each resource has a specific configuration that determines the number (N) of cores that can be dedicated to a job without slowing the code (and other user and system codes). Each Resource Provider (RP) defines for its system the minimum number of SUs charged for a job running in the default batch queue, calculated as wallclock runtime multiplied by N. For the minimum charges on TeraGrid resources that use this calculation, see the table below.
Note: The actual charge will depend on the
specific requirements of your job (e.g., the mapping of the cores
across the machine, or the priority you wish to obtain). Consult each
system's user guide for details. If you have questions, contact
help@teragrid.org .
| Resource | Minimum charge in SUs |
|---|---|
| Ember (NCSA) | wallclock time * 6 cores (see additional information below) |
| Kraken (NICS) | wallclock time * 12 cores |
| Blacklight (PSC) | wallclock time * 16 cores |
| Pople (PSC) | wallclock time * 4 cores |
| Condor Pool (Purdue) | CPU hours used (see additional information below) |
| Steele (Purdue) | wallclock time * number of CPUs
requested (see additional information below) |
| Dash (SDSC) | batch queue = wallclock time * 8
cores large queue = wallclock time (see additional information below) |
| Trestles (SDSC) | normal queue = wallclock time *
32 cores shared queue = wallclock time (see additional information below) |
| Lonestar (TACC) | wallclock time * 12 cores (see additional information below) |
| Ranger (TACC) | wallclock time * 16 cores (see additional information below) |
| Spur (TACC) | wallclock time * 16 cores |
Additional information
Ember (NCSA): Jobs are allocated the resources required to cover their request (cores and memory) in multiples of six.
Condor Pool (Purdue): It's impossible to adjust for CPU speed, but CPU speeds do not vary greatly across the pool.
Steele (Purdue): Charges are calculated at [walltime * CPUs requested] unless exclusive access is requested, in which case the calculation is [walltime used * nodes requested * 8] (eight being the number of CPUs in Steele's nodes).
Dash (SDSC): On regular compute nodes (batch queue), you are charged for a full node even if you use less than eight cores. The vSMP node is charged based on the fraction of node used based on the max (fraction of cores used, fraction of memory used).
Trestles (SDSC): For the normal queue, you are charged for a full node even if you use less than 32 cores. For the shared queue, you are not charged for a full node, and the node is shared with other users/jobs.
Lonestar (TACC): You are charged for the entire blade, even if you use less than 12 cores.
Ranger (TACC): You are charged for the entire blade, even if you use less than 16 cores.
Last modified on January 20, 2012.
