Using HBase PerformanceEvaluation (PE) Tool for Read/Write Latency Benchmarking (P99/P999)

When using HBase, it is essential to understand server performance; therefore a baseline latency test (P99/P999) is recommended before deployment.

The article introduces HBase's built‑in performance testing tool org.apache.hadoop.hbase.PerformanceEvaluation (PE) and shows how to invoke it with bin/hbase pe together with various options such as --nomapred , --oneCon , --rows , --size , --valueSize , --compress , --presplit , --autoFlush , and others.

Key parameters are described, including how to control threading, connection sharing, value size, table name, number of rows, data size, compression, and region pre‑splitting.

Four benchmark cases are executed on a cluster (2 HMasters, 8 RegionServers) using 16 client threads: randomWrite, sequentialWrite, randomRead, and sequentialRead. Sample command lines and the resulting latency statistics (mean, min, max, 50th‑99.999th percentiles) are presented.

Results show random write P999 latency around 4 ms, sequential write around 1.2 ms, random read around 78 ms, and sequential read around 75 ms, demonstrating the tool’s ability to quantify HBase read/write performance.

The article concludes that PE is useful for measuring HBase latency (P99/P999) but does not cover throughput metrics such as TPS, which require tools like YCSB.