Difference between revisions of "Main Page"
m |
m (→Job priorities) |
||
| Line 58: | Line 58: | ||
At any given time every queue is only allowed a maximum number of cores to ensure quick jobs aren't unnecessarily blocked by long running jobs. | At any given time every queue is only allowed a maximum number of cores to ensure quick jobs aren't unnecessarily blocked by long running jobs. | ||
| − | * '''short''' - | + | * '''short''' - 150 jobs per person, highest priority |
| − | * '''day''' - | + | * '''day''' - 150 jobs per person |
* '''week''' - 800 cores, burstable to 1000 if cluster is idle | * '''week''' - 800 cores, burstable to 1000 if cluster is idle | ||
* '''month''' - 500 cores (burstable to 600), maximum of 10 jobs per user (burstable to 20), maximum of 200 cores per user (burstable to 300) | * '''month''' - 500 cores (burstable to 600), maximum of 10 jobs per user (burstable to 20), maximum of 200 cores per user (burstable to 300) | ||
Revision as of 10:29, 24 March 2017
| HPC1 will be down for maintenance from April 3rd to April 7th 2017. |
The University of Stellenbosch hosts multiple HPCs (High Performance Computing clusters). This wiki provides information on the two largest systems, HPC1 and CAF-HPC1.
Contents
HPC1 (also known as Rhasatsha)
HPC1 is available to all users registered on campus. In essence, if you have a network login, you can use this HPC.
All users are granted 1000 CPU hours to test the system and determine its usefulness. Once the 1000 hour quota is depleted, users are required to pay a registration fee to gain unlimited access.
Free users are granted 1000 CPU hours and a 10GB disk quota. Paid users are granted unlimited CPU and a 1TB disk quota.
See HOWTO register for details on how to register.
General information
Feel free to contact Charl Möller (x9490) with any queries regarding the cluster.
rhasatsha: (the rha is pronounced as gaan in Afrikaans) a clever person/object; highly intelligent; something that acts promptly; a wide awake person/object who/that is always on the spot; a versatile person/object that can tackle anything successfully.
Monitoring tools
- Ganglia cluster monitor (only available on campus)
- Munin health monitor (only available on campus)
- XDMoD usage explorer (only available on campus)
Specifications
The HPC currently has the following compute specifications:
- 11x 8-core Intel Xeon E5440 @ 2.83GHz with 16GB RAM
- 17x 48-core AMD Opteron 6172 @ 2.10GHz with 96GB RAM, Infiniband interconnect
- 2x 64-core AMD Opteron 6274 @ 2.20GHz with 128GB RAM, Infiniband interconnect
- 8x 8-core Intel Xeon X5450 @ 3.00GHz with 32GB RAM
- 2x 8-core Intel Xeon X5450 @ 3.00GHz with 24GB RAM
- 1x 64-core AMD Opteron 6366 HE @ 1.8GHz with 128GB RAM, Infiniband interconnect
- 2x 48-core Intel Xeon E5-2670 v3 @ 2.30GHz with 512GB RAM, Infiniband interconnect
The total is 1272 available cores.
Job priorities
The HPC currently has 5 queues into which jobs are automatically divided based on walltime requested.
- short - queue for jobs running up to 2 hours (#PBS -l walltime=2:00:00)
- day - queue for jobs running up to 24 hours (#PBS -l walltime=24:00:00)
- week - queue for jobs running up to 7 days (#PBS -l walltime=168:00:00)
- month - queue for jobs running up to 31 days (#PBS -l walltime=744:00:00)
- long - queue for jobs running longer than 31 days
At any given time every queue is only allowed a maximum number of cores to ensure quick jobs aren't unnecessarily blocked by long running jobs.
- short - 150 jobs per person, highest priority
- day - 150 jobs per person
- week - 800 cores, burstable to 1000 if cluster is idle
- month - 500 cores (burstable to 600), maximum of 10 jobs per user (burstable to 20), maximum of 200 cores per user (burstable to 300)
- long - 400 cores (burstable to 600), maximum of 10 jobs per user (burstable to 20), maximum of 200 cores per user (burstable to 300)
It is imperative that you accurately estimate your job's running time. Estimate too high, and you may find yourself in an unfavourable queue. Estimate too low and the job will be killed by the system when the walltime is reached. Once a job is running, only the administrator can increase the walltime. All jobs are required to specify a walltime.
Furthermore, all interactive jobs will be satisfied by the test queue and will be limited to a maximum of 8 cores and walltime of 24 hours.
Acceptable usage
This system may only be used for bona fide academic work. Any effort to use it for consultancy work or any other commercial purpose may lead to the permanent banning of the user from the system.
Citations
We require an acknowledgement in any thesis, paper, publication or presentation that references results computed on this system. In addition we would like to be able to reference these published works.
Suggested form of acknowledgement:
Computations were performed using the University of Stellenbosch's HPC1 (Rhasatsha): http://www.sun.ac.za/hpc
CAF-HPC1
CAF-HPC1 is only available to registered users.
See HOWTO register for details on how to register.
General information
Feel free to contact Carel van Heerden (x9044) with any queries regarding the cluster.
Monitoring tools
- Ganglia Cluster Monitor (only available on campus)
- Munin health monitor (only available on campus)
Specifications
The HPC currently has the following compute specifications:
- 1x 80-core Intel Xeon E7-4850 @ 2.00GHz with 1024GB RAM, Infiniband interconnect
- 3x 48-core Intel Xeon E5-2650 v4 @ 2.20GHz with 512GB RAM, Infiniband interconnect
- 2x 64-core AMD Opteron 6274 @ 2.20GHz with 128GB RAM, Infiniband interconnect
- 2x 24-core Intel Xeon X5650 @ 2.67GHz with 48GB RAM, Infiniband interconnect
- 3x 16-core Intel Xeon E5530 @ 2.40GHz with 24GB RAM
The total is 448 available cores.
Job priorities
The HPC currently has 5 queues into which jobs are automatically divided based on walltime requested.
- short - queue for jobs running up to 2 hours (#PBS -l walltime=2:00:00)
- day - queue for jobs running up to 24 hours (#PBS -l walltime=24:00:00)
- week - queue for jobs running up to 7 days (#PBS -l walltime=168:00:00)
- month - queue for jobs running up to 31 days (#PBS -l walltime=744:00:00)
- long - queue for jobs running longer than 31 days
At any given time every queue is only allowed a maximum number of cores to ensure quick jobs aren't unnecessarily blocked by long running jobs.
- short - unlimited, highest priority
- day - unlimited
- week - 300 cores, burstable to 400 if cluster is idle
- month - 200 cores (burstable to 300), maximum of 3 jobs per user (burstable to 5), maximum of 100 cores per user (burstable to 200)
- long - 100 cores (burstable to 200), maximum of 3 jobs per user (burstable to 5), maximum of 50 cores per user (burstable to 100)
It is imperative that you accurately estimate your job's running time. Estimate too high, and you may find yourself in an unfavourable queue. Estimate too low and the job will be killed by the system when the walltime is reached. Once a job is running, only the administrator can increase the walltime.
Any job that does not specify a walltime will be assigned a default of 5 minutes.
