Grasper: A High Performance Distributed System for OLAP on Property Graphs
Graph analytics has a broad spectrum of applications in both academia and industry. However, processing online graph analytical queries (or OLAP) remains to be challenging since it has much stricter requirements on both latency and throughput. Existing systems suffer from various performance problems due to the key challenging factors as follows:
- Diverse query complexity: online analytical queries may differ signifcantly in their workloads (e.g., from a simple property check on a vertex to complicated pattern matching), and thus a mechanism is needed to support high parallelism and load balancing for heavy-workload queries, while at the same time preventing the starvation of light workload queries.
- Diverse data access patterns: query operators (e.g., flter, traversal, count) often have diverse access patterns on data and hence require different optimizations (e.g., cache, index), which makes it challenging to design a unifed computational model that optimizes the execution of different query operators.
- High communication and CPU costs: a query may have complex execution logics such as graph traversals that access a large portion of a PG, aggregation that collects intermediate results to one place through network, and joins that are CPU- and data-intensive. Such complex logics often result in high overheads on both communication and computation.
We propose Grasper, a distributed system designed to address the aforementioned challenges of processing online analytical queries on PGs. Grasper adopts Remote Direct Memory Access (RDMA) to reduce the high network communication cost for distributed query processing and introduces an RDMA-enabled native PG storage to exploit the benefts of RDMA. The key design of Grasper is a novel query execution model, called Expert Model, which achieves high utilization of CPU and network resources to maximize the processing effciency. There are three core benefts brought by Expert Model:
- It allows Grasper to support high concurrency in processing multiple queries and adaptive parallelism control within each query according to its workload.
- It enables tailored optimizations for different categories of query operations based on their own data access patterns.
- Underlying system optimizations such as memory locality and thread-level load balancing are also incorporated into the design, which are critical for achieving millisecond latency for processing complex analytical queries.
You can use git clone or just download zip archive to get the codes. The source code of Grasper is available here