Clark School
This Site

HPC Large Allocation Proposals

This form is used to request access to Engineering specific cycles on DeepThought2 (DT2). These proposals will be evaluated by an Engineering faculty committee on an annual basis with award dates on the first day of October. Engineering has approximately 20M Service Units/Year to allocate in this fashion. A service unit (SU) is 1 Core Hour (most nodes on DT2 have two 10 core nodes = 20 cores per node). 

All Tenure/Tenure-Track faculty will be granted a start-up allocation of 50K SU's through a separate application. These cycles are granted as soon as possible without a faculty review. This Engineering HPC allocation proposal is for requesting time beyond a start-up allocation and will be awarded for a one year period (equally divided over each quarter). 

The campus also has a separate application process that our faculty are eligible to request cycles. These allocations all run on the same DT2 infrastructure


Scalability for HPC

The figure shows the parallel performance of 2 different solvers (left plot and right plot). For each solver, the parallel performance is measured by a speedup factor S_N (left y-axis) and an efficiency factor E_N (right y-axis) as a function of number N of processing elements (PEs, meaning cores) actually used (x-axis). 

Parallel speedup S_N = (time to compute problem using 1 PE) / (time to compute same problem using N PEs) 

Parallel efficiency E_N = (time to compute problem using 1 PE) / (N x (time to compute same problem using N PEs)) = (S_N/N) 

Ideally, S_N = N and E_N = 1. Values of S_N below N and E_N below 1 indicate a loss of parallel performance. Note that this test has to be done for a problem of sufficient size and relevant-to/representative-of the user application area.