bloomen / gcl Goto Github PK

To improve performance we want to investigate using a custom mem allocator for tasks. In particular, an arena allocator may be useful.

Apparently, using futures is not the most efficient approach so we should investigate whether we can have a dedicated scheduler which collects all tasks and then offloads them to workers. Each worker could have its own queue. Once a task's parents have all finished a task is marked as ready and given to any worker. The worker will assign the result once finished.

Suggestion

I noticed that the cache entries listed in CMakeLists.txt are all-lowercased.

However according to my observation, they are usually all-uppercased. At least, letters except the package name should be all-uppercased. That is:

• Scheme 1:

  • GCL_BUILD_TESTS
  • GCL_ENABLE_ASAN
  • GCL_ENABLE_TSAN
  • GCL_ENABLE_COVERAGE

• Scheme 2:

  • gcl_BUILD_TESTS
  • gcl_ENABLE_ASAN
  • gcl_ENABLE_TSAN
  • gcl_ENABLE_COVERAGE

Versions

• gcl: 1.1.0

Currently, gcl uses the simplest static task scheduling there is which may unnecessarily block worker threads. This may be useful: http://charm.cs.uiuc.edu/users/arya/docs/6.pdf

Can tasks_by_breadth be made to run faster? Surely...

We want to use the address and thread sanitizers when running CI builds.