Main goals

• Easy compose of any-size blocks as Lego
• "This was last time I wrote function that do X in env Y"
  • Connect MY service A to MY service B. Sync protocols. Ensure that all works.
  • Decide where code should be executed. Protocol. Start/restart/stop. Coordinator. Multithread management. Priority. Data locality. CPU cache hidden miss. Perf manual tuning.
  • Write OpenCL kernel
• Single machine is not enough. Multithreading must have but it's only beginning.
• Single GPU is not enough. Managing multiple kernels on multiple GPU must be easy.
• 2+? level design.
  • Top level pipe scheme
  • May be some intermediate level
  • Low level implementation of blocks
• "Write once debug everywhere" is not bad. At last you "write once".
• 1-space change semantics should be avoided

a +b
a + b
a+b

a?b:c
a? b:c
a ? b:c

• Sync and async operations on top level design make no difference.
  • Let compiler decides how to implement it.
  • You can fix it in monitoring.
  • [LATER] You can fix it in special annotations.