Create a model to evaluate CPU and GPU performance dynamically for metadirective branching about llnl-work-notes

[ouankou]

FPM (functional performance modeling) is used to create a model for offloading guidance.
In general, the computing speed of a given device s = f(x), where x is problem size. f(x) is application-specific and it may not be a linear function.

2D Stencil running on Lassen:
Problem size: 64
Iterations: 5000
CPU total time: 1275 ms
GPU total time: 836 ms

Problem size: 32
Iterations: 5000
CPU total time: 456 ms
GPU total time: 827 ms

[ouankou]

The scope of profiling should be limited by user-defined parameter. The problem size should be provided by users per application.
For example, if the problem size varies between 16 to 64, then there's no point to profile 1024. Otherwise, the overhead of unnecessary profiling is too much for this particular application.

[ouankou]

The description of FPM has been added to Overleaf.

A complete FPM-based model is supposed to be created offline in advance. During runtime, a partial model based on selected problem sizes can be created to represent the complete model.

[ouankou]

For 2D stencil on Pascal:
CPU: 72 cores, GPU: NVIDIA Tesla P100 16 GB.

The CPU performance model (16 threads) is: y = -19521 + 8628 * ln(x), where x is the problem size and y is how many elements are processed per ms.

The GPU performance model (Tesla P100) is: y = -54023 + 16517 * ln(x)

The model data and equation are on Google Sheet.

[ouankou]

For 2D stencil on Surface:
CPU: 32 cores, GPU: NVIDIA Tesla K40 12 GB.

The CPU performance model (8 threads) is: y = -2576 + 3610 * ln(x), where x is the problem size and y is how many elements are processed per ms.

The GPU performance model (Tesla K40) is: y = -20039 + 7341 * ln(x)

The model data and equation are on Google Sheet.