Tuple Graph Subgraph Partitioning for Distributed Computing Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current programming models for cloud computing face challenges in achieving efficient distributed computation due to performance issues related to data proximity and transfer costs, particularly in executing distributed low-latency non-batch programs.
Innovation Solution
A new programming model that constructs a program as a directed acyclic graph of operators, allowing operators to run on different machines, with automatic location assignment and graph partitioning to minimize network traffic and optimize execution, using techniques like sharding and pipelining to enhance latency and data efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If operators are distributed across different machines in cloud computing, then computing efficiency and resource utilization are improved, but network traffic and data transfer costs increase
Solution Approach 1:
The patent segments the computation graph into multiple subgraphs and distributes them across different machines. Each subgraph is executed independently on a separate machine, allowing parallel processing while minimizing cross-machine data transfer. The segmentation is performed in two phases: main partitioning that divides the overall graph, and local partitioning that further optimizes each subgraph.
Solution Approach 2:
The patent applies local quality optimization by performing local partitioning on each subgraph after main partitioning. This ensures that operations within each subgraph are optimized for their specific machine context, reducing local data transfer requirements and improving computing efficiency on each individual machine.
2Productivity
If data is transferred between computation units in distributed systems, then processing capability is improved, but latency increases
Solution Approach 1:
The patent performs preliminary actions by pre-assigning locations to operations during the partitioning phase, before actual execution. The main partitioning algorithm determines the optimal machine for each subgraph in advance, and this assignment is stored and used during execution. This eliminates the need for dynamic location decisions during runtime, reducing latency.
Solution Approach 2:
The patent introduces an intermediary partitioning algorithm that acts as a mediator between the computation graph and the distributed system. This algorithm determines the optimal distribution of subgraphs across machines, balancing processing capability requirements with latency minimization by considering data locality and communication patterns.
3Productivity
If the computation graph is partitioned into more subgraphs, then parallel execution is improved, but the complexity of location assignment increases
Solution Approach 1:
The patent segments the location assignment process into two distinct phases: main partitioning that divides the overall graph into subgraphs and assigns them to different machines, and local partitioning that optimizes each subgraph independently. This segmentation of the assignment process reduces overall complexity compared to a single monolithic partitioning approach.
Solution Approach 2:
The patent applies partial action by performing location assignment in stages rather than all at once. The main partitioning phase assigns subgraphs to machines, and then local partitioning phases refine assignments within each subgraph. This staged approach makes the complex location assignment problem more manageable and computationally feasible.
Data Source
Figure 1
Figure 2A~2B
Figure 3
AI summary
A programming model generates a graph for a program, the graph including a plurality of nodes and edges, wherein each node of the graph represents an operation and edges between the nodes represent streams of data input to and output from the operations represented by the nodes. The model determines where in a distributed architecture to execute the operations represented by the nodes. Such determining may include determining which nodes have location restrictions, assigning locations to each node having a location restriction based on the restriction, and partitioning the graph into a plurality of subgraphs, the partitioning including assigning locations to nodes without location restrictions in accordance with a first set of constraints, wherein each node within a particular subgraph is assigned to the same location. Each of the subgraphs is executed at its assigned location in a respective single thread.