Query serving benchmarks: sysbench results

​

Performance highlights

• Low latency: RisingWave consistently delivers low-latency query responses, with average latencies as low as 4.96ms for point selects and 13.84ms - 16.40ms for more complex range queries. P95 latencies are also within a desirable range, indicating predictable performance for most requests.
• High throughput: RisingWave achieves high query throughput, processing up to 25,814 queries per second (QPS) for point selects and over 8,000 QPS for random point and range selects.
• Stable performance: RisingWave exhibits stable performance even when the data size increases from 1 million rows to 10 million rows, as long as the data can be cached in memory.

​

Benchmark methodology

​

Environment

• Hardware:
  • One AWS EC2 instance (8 vCPUs, 16GB memory) for the RisingWave compute node.
  • One AWS EC2 instance (8 vCPUs, 16GB memory) for the RisingWave frontend node.
  • One AWS EC2 instance (8 vCPUs, 16GB memory) for the sysbench client (query generator).
• Software: RisingWave (specific version not mentioned in the report, but presumably a recent stable or nightly build).

​

Test procedure

1. A table named sbtest was created in RisingWave with the following schema:
   Copy

   Ask AI

   CREATE TABLE sbtest(
     id INTEGER,
     k INTEGER,
     c VARCHAR,
     pad VARCHAR,
     PRIMARY KEY (id)
   );
   

2. Data was inserted into the sbtest table. Two data sizes were tested: 1 million rows (approximately 186MB) and 10 million rows (approximately 1.86GB).
3. 128 threads were used to concurrently issue queries against the table. This level of concurrency is sufficient to saturate the CPU of the RisingWave deployment.
4. Results, including latency and throughput, were collected.

​

Workload

• oltp_point_select:
  Copy

  Ask AI

  SELECT c FROM sbtest WHERE id = ?;
  

• select_random_points:
  Copy

  Ask AI

  SELECT id, k, c, pad FROM sbtest WHERE k IN (?, ?, ?, ?, ?, ?, ?, ?, ?, ?);
  

• select_random_ranges:
  Copy

  Ask AI

  SELECT count(k) FROM sbtest WHERE k BETWEEN ? AND ? OR k BETWEEN ? AND ?;
  

​

Detailed benchmark results

​

Key observations

• Memory impact: When the memory is sufficient to cache all the queried data, performance remains relatively stable as the table size grows. This highlights the importance of memory capacity for OLTP workloads.
• Bottlenecks: In the oltp_point_select workload, the frontend node becomes the primary bottleneck due to the lightweight nature of the operations. For select_random_points and select_random_ranges, both the compute and frontend nodes experience CPU saturation.