filprofiler

filprofiler copied to clipboard

Using runtime instrumentation would allow not having to start with Fil from the very start, and would make it safe to use with production servers. In particular, for long-running servers...

itamarst

Investigate GPU memory profiling

1

For CUDA it's possible we can track allocations by intercepting `cudaMalloc()` and friends. This would need to be tracked and reported separately than CPU memory. https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__MEMORY.html

itamarst

Once #12 is done, consider support for Spyrder as another scientific computing environment.

itamarst

Users might want to parse that, and the `@@TB` stuff is internal implementation detail that shouldn't be leaked.

itamarst

For profiling, the real usage pattern is: 1. Run with current code. 2. Try to fix code. 3. Run again, figure out difference, go to step 2 if not fixed....

itamarst

itamarst

Play around with key/value databases

1

### Goals: 1. Persistence could enable better UX, e.g. for crashes. 2. Reduce memory overhead from tracking allocations. ### Things to look for: 1. Ability to create snapshots, for peak...

itamarst

Would need to compare performance to the native allocator first.

itamarst

Try to reduce memory usage by using nested data structures

1

The theory: memory tracking overhead mostly matters if you have lots of small allocations. If you have lots of small allocations, they will end up in similar parts of the...

itamarst

Additionally, replace line number with bytecode index, deferring line number calculation to report-generation time. Together these changes should speed up the tracing part of the proflier. This in progress in...

itamarst