Threading Models/Multicore Programming

Objectives

1. Understand how threads relate to parallel programming
2. Understand models considered for implementing threads

Notes

• Concurrency: allowing all the tasks on a system to make progress
  • Interrupt based scheduling as we have discussed
  • Only requires one cpu/core/...
• Parallelism: allowing multiple process to run at the same time
  • Requires multiple cores/cpu/...
• The book identifies five challenges with parallel programming
  • Identifying independent tasks.
  • Balance: Keeping all processors busy
  • Division of data: can you get the data to the right place at the right time, and share the data you need.
  • Data dependency: can you keep the data synchronized, with the correct process modifying/reading it at the correct time.
  • Testing and debugging: Parallel introduces race conditions, dead lock, and non-determinism. More in the next chapter (I think)
• Types of parallelism
  • Data parallelism: SPMD
    • Split the data across workers.
    • Each worker does the same thing.
    • This, IMHO, is the easiest, and what I know how to do.
  • Task parallelism MPMD
    • Break the problem into tasks
    • Each does something different.
    • Coordinate the workers
    • This is hard/easy, depending on the task.
  Multithreading models
  • Look at the pictures on 166-167
  • User threads - run in the process
  • Kernel threads - run in the kernel
  • Many to one
    • The process handles threads, the kernel knows nothing about them.
    • The thread library must handle creation, scheduling, destruction, ...
    • Does not take advantage of multiple cores/cpus
    • The original threading packages.
    • Probably very few applications any longer.
  • One to one
    • Each thread in a process is supported by a thread in the kernel.
    • May overload the kernel with too many threads.
    • Probably the most common. (windows, linux, Unix, ...)
  • Many to many
    • Multiple user threads may map to a single kernel thread
    • Perhaps fewer kernel threads
    • The kernel less likely to become overloaded
    • but "Although the many-to many model appears to be the most flexible ... in practice it is difficult to implement"
    • And as cores increase this becomes less important.