Scheduling Dataflow Components via Latency Modeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing distributed computing platforms face challenges in guaranteeing soft-real-time requirements due to difficulties in measuring resource needs and dynamic partitioning, leading to unpredictable behavior under varying processor utilization without over-provisioning.
Innovation Solution
A method and device for scheduling dataflow components that determine the maximum serving rate by modeling total latency as a sum of queuing and processing latency in a D/M/1 queuing system, allowing for the selection of processing devices based on resource availability and latency, thereby limiting queuing latency to ensure soft-real-time requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If known scheduling solutions are used that assume known resource needs, then scheduling simplicity is improved, but measurement accuracy deteriorates because resource needs are hard to measure in distributed systems with dynamic partitioning
Solution Approach 1:
The system performs self-characterization by automatically determining the maximum serving rate of dataflow components through controlled message flows and latency measurements, eliminating the need for external resource need assessments or intrusive instrumentation of processing nodes
Solution Approach 2:
The invention changes the approach from assuming known resource parameters to dynamically determining serving rate parameters through systematic measurement. By varying message flows and measuring latencies, the system derives accurate serving rate parameters that reflect actual system behavior under different loads
2Reliability
If over-provisioning is used to guarantee soft-real-time requirements, then reliability is improved, but productivity deteriorates due to wasted processing resources
Solution Approach 1:
The system establishes feedback loops that continuously monitor actual serving rates and latencies, using this information to dynamically adjust scheduling decisions. This feedback mechanism ensures soft-real-time requirements are met while optimizing resource utilization, eliminating the need for conservative over-provisioning
Solution Approach 2:
The invention transitions from static scheduling assumptions to dynamic serving rate determination. By continuously characterizing the actual serving rate of dataflow components under varying loads, the system adapts scheduling decisions to real-time conditions, ensuring reliability without wasted resources
3Measurement precision
If intrusive instrumentation is used to measure resource needs, then measurement accuracy is improved, but device complexity deteriorates
Solution Approach 1:
The invention introduces a scheduling device as an intermediary that performs non-intrusive measurements by injecting controlled message flows through the databroker layer. This intermediary approach enables accurate serving rate characterization without requiring instrumentation of processing nodes or modification of dataflow components
Solution Approach 2:
The measurement functionality is extracted from the processing nodes and consolidated in the scheduling device. By taking out the instrumentation requirements from the distributed processing system and centralizing them in the scheduler, the invention achieves accurate measurements while reducing overall system complexity
Data Source
Figure 1~3
AI summary
Embodiments relate to a method for scheduling dataflow components (6), executed by a scheduling device (4), comprising repeating the following steps for successive increasing input message rates (1/β): - applying (S1) an ingress message flow (F1) at an ingress message rate (1/β) to a distributed dataflow comprising at least one dataflow component (6) executed by respective processing devices (5), - determining (52) a total latency (Δt) of at least one of dataflow component (6), and - determining (53) a serving rate (µ) associated with said ingress message rate (1/β) for said at least one dataflow component (6), in function of said ingress message rate (1/β) and said total latency (Δt), until the scheduling device (4) is unable to determine a serving rate (µ) associated with the ingress message rate (1/β), thereby determining a maximum serving rate for said at least one dataflow component (6).