Load balancing for parallel tasks
View Patent ↗Techniques for balancing processing loads when performing parallel tasks on one or more processing nodes that share memory resources are provided. For some embodiments, the techniques involve distributing work according to an alternate block cyclic distribution scheme of at least one dimension, wherein work processes are assigned to sets of threads in at least first and second adjacent blocks based on round-robin manner according to a first sequence and an alternate round-robin manner according to a second sequence that is a mirror image of the first sequence, respectively.
1. A method of distributing a plurality of computational workloads between a plurality of processing threads executing in a parallel computing system, the method comprising:
arranging, by a processing node in the parallel computing system, the plurality of computational workloads in order of computational cost;
dividing the arranged computational workloads into a plurality of work cycles, each work cycle having a number of arranged computational workloads equal to a multiple of the number of the plurality of processing threads; and
distributing the computational workloads among the plurality of processing threads according to a sequence of distribution patterns, wherein:
a first work cycle distributes its plurality of computational workloads to the plurality of processing threads according to a first distribution pattern in the sequence, and
each subsequent work cycle distributes its plurality of computational workloads to the plurality of processing threads according to a distribution pattern that is the reverse of the distribution pattern immediately preceding it in the sequence,
wherein a processing thread that receives a high cost workload in a given work cycle receives a low cost workload in a subsequent work cycle, thereby balancing the plurality of computational workloads among the plurality of processing threads.
2. The method of claim 1 , wherein each work cycle consists of one workload for each processing thread, so that the workloads are distributed according to an alternating simple cyclic scheme.
3. The method of claim 1 , wherein each work cycle consists of the same number of at least two workloads for each processing thread, so that the workloads are distributed according to an alternating block cyclic scheme.
4. The method of claim 1 , wherein the threads are provided with a plurality of orderings, the first work cycle distributes the first set of workloads according to a distribution pattern based on the plurality of orderings, and the second work cycle distributes the second set of workloads according to a distribution pattern that is the reverse of at least one of the plurality of orderings.
5. The method of claim 1 , wherein the workloads relate to an application selected from the group consisting of: multiplying a matrix and a vector, factoring a matrix, and solving a linear system of equations.