|This page to be deleted.|
This page provides documentation for all of SHARCNET's computational resources. This includes both hardware specifications as well as specialized usage information as appropriate, and links to information in the web portal.
Which system should I use?
Conventional HPC computing should be done on the core systems based on the particular memory and scaling requirements for the given program. Here is the snapshot with major systems in SHARCNET.
As a rough suggestion:
- serial programs should be run on kraken
- threaded programs (shared-memory) should run on saw, kraken or possibly orca, if they can scale well
- mpi programs (distributed-memory) should run on requin, saw or orca
- large memory jobs can be run in orca (note: hound system is no longer available)
- the MPI jobs needs to run with high network bandwidth may got all core systems
- the jobs require high disk I/O may go to orca/saw/requin.
If one has special computing needs (acceleration, visualization, storage, etc.) please see the specialty systems.
There are also a wide array of contributed systems where one may find their jobs run faster than on the core systems and users are invited to try them as well, though they are subject to availability depending on the system's particular scheduling arrangement with the contributor.
This table indicates the primary intent of the system and includes a comparison of common system specifications in terms of their relative performance.
The legend is as follows:
|Core Systems||Target Utilization||core/node||CPU clock Hz||FLOPS/cycle||MEM Size/node||MEM Bandwidth/socket||Cache Size/node||Interconnect||Parallel|
|orca||large MPI capacity cluster|
|saw||large MPI capacity cluster|
GPU accelerated Systems
Please check Storage_systems.
|Visualization Workstations||Graphical programs|
|Specialty Systems||Target Utilization|
|iqaluk||SMP system, fast storage|
|redfin||large-memory compute cluster|
How long will my job wait in the queue before it starts to run?
Please see the Current wait time distribution page to see how long various jobs of different sizes have taken to run on each of the core systems. Depending on how busy the systems are you may want to avoid them if you are looking for the fastest time to results.
One can also find important statistics about jobs being scheduled on the clusters in the web portal cluster performance page.