When a tensor graph is submitted into Mars scheduler, a graph comprises of
operands and chunks will be generated given
chunk_size parameters passed in
After tiling a tensor graph into a chunk graph, we will combine adjacent nodes
to reduce graph size as well as to utilize acceleration libraries such as
numexpr. Currently Mars only merges operands that forms a single chain
without branches. For example, when executing code
import mars.tensor as mt a = mt.random.rand(100, chunk_size=100) b = mt.random.rand(100, chunk_size=100) c = (a + b).sum()
Mars will compose operand ADD and SUM into one FUSE node. RAND operands are excluded because they don’t form a line with ADD and SUM.
Initial Worker Assignment¶
Assigning operands to workers are crucial to the performance of graph execution. Random worker assignment will contribute to huge network cost and imbalanced workload between different workers. Since the workers of non-initial operands can be effectively decided given data distribution and cluster idleness, we only assign workers for initial nodes in graph preparation stage.
Initial worker assignment should obey several principles. First, the number of operands assigned to each worker should be balanced. This makes full use of the cluster especially in the late stage of graph execution. Secondly, operand assignment should minify the amount of network transfer in its descendants. That is, locality need to be observed in the assignment process.
Note that these principles sometimes collides with each other. That is, a network-minimal solution may be quite biased. We developed a heuristic algorithm in practice that takes a balance between minimal network transfer and worker load balance. The algorithm is described below:
Select the first worker who does not have any operands;
Start breadth-first search on the undirected graph produced from the operand graph;
When an initial operand is visited, we assign it to the worker we selected in Step 1;
Stop assignment when the number of operands visited is greater than the average number of operands for every worker;
Go to Step 1 when there are workers left.