Towards Framework-Independent, Non-Intrusive Performance Characterization for Dataflow Computation

Grade10: A Framework for Performance Characterization of Distributed Graph Processing

“…Software resources include locks and queues, or even runtime services such as a garbage collector (GC), a data store, or a synchronization service. [11] ~~…”

“…In comparison, Grade10 offers more fine-grained resources utilization by upsamling to individual execution phases for single-workload bottleneck detection. Closest to Grade10 is work from Tian et al [11] which also does performance characterization using a DAG-based computation model with system-level resource monitoring. However, Grade10 captures a more comprehensive set of performance issues (e.g., including burstiness, imbalance), does more fine-grained attribution across time and execution phases (in comparison to coarser machine learning based attribution used by Tian et al), and is more thoroughly evaluated with two state-of-the-art graph frameworks, 2 datasets, and 4 algorithms (than just two workloads by Tian et al).…”

“…Although many distributed graph processing frameworks already exist [4]- [8], they need to adapt to these trends and process graphs increasingly faster. There are many timely performance challenges specific to graph processinge.g., irregular workloads, idiosyncratic operation of graph processing frameworks-, which generic performance characterization frameworks [9]- [11] cannot fully address. In contrast, in this work we propose Grade10, a specialized performance characterization framework for distributed graph processing.…”

“…Performance characterization and analysis have a long and successful tradition. Generic tools and frameworks [9]- [11] are helpful in identifying performance issues, scalability bottlenecks, and execution inefficiencies. However, performance analysis and characterization of a distributed graph processing framework remains daunting.…”