Home/ Questions/Q 8174675
In Process

The Archive Base Latest Questions

Editorial Team
  • 0
Asked: 2026-06-06T22:31:21+00:00

The algorithm is described in this paper: Thread Scheduling for Multiprogrammed Multiprocessors . Briefly,

  • 0

The algorithm is described in this paper: Thread Scheduling for Multiprogrammed Multiprocessors. Briefly, a computation is distrubuted in processes and each one has a deque of threads to do the job. A process can push (pop) threads to (from) the bottom of its deque and other processes can work steal from it by popping threads from the top. Thus, work can be dinamically created by the push operation. The algorithm is the following.

Schedulling Algorithm

My question is about the popTop() work stealling function. I don’t think it will work properly for all cases. For example, suppose a process A which has its queue Q and a process B that is trying to steal work from Q, calling popTop(). Suppose also that B is preempted after line 2 of popTop() and localBot = X at this moment. If A runs and popBottom() until the bottom of Q <= X, when B resumes its run it will get a thread that already have been processed by A.

Are my thoughts correct? I need to verify it because I will implement it to do work balancing in a CUDA program.

Share

Leave an answer

You must login to add an answer.

Need An Account,

1 Answer

  1. Editorial Team

    The code is using cas() (compare and swap) to try and stop the sort of thing you are describing. If popTop() stops after line 2, it stops after having read in age and bot. If popBottom() then runs and returns a thread, it will have incremented fields within age and written the incremented version back into memory using cas(). Now when B resumes and calls cas() the cas() instruction finds that B the values that B has provided for age do not match the values in memory (which means that this call to cas() does not modify memory). So B finds that (oldAge == newAge) and returns ABORT. In these circumstances you would normally try again and hope for better luck next time. The article seems to be saying that calls to yield() are necessary for you to have decent luck, but in any case popTop() should not return a thread somebody else has grabbed.

    There is of course a Wikipedia article on cas() at http://en.wikipedia.org/wiki/Compare-and-swap.

    I would place parallel code using locks one level of difficulty above serial code, and lock-free parallel code one level of difficulty above locking parallel code. I would not write lock-free parallel code unless I knew for sure that I needed performance, and there was no existing known trustworthy code that I could reuse. I would not trust such code until I had tested it exhaustively, and I would actually prefer to have model-checking if possible as well.

    • 0