ClickHouse Managed Postgres 团队介绍了一种优化方案，旨在解决单线程 PgBouncer 在多核 CPU 环境下的性能瓶颈。该服务默认采用单线程架构运行，在配备 16 个虚拟核心的机器上仅占用一个核心，导致其余计算资源闲置 [1]。为充分利用服务器硬件能力，ClickHouse Managed Postgres 部署了与可用核心数量成比例的 PgBouncer 进程集群，并将所有进程绑定至同一端口以启用 Linux 内核的 so_reuseport 特性 [1]。
针对因内核负载均衡机制可能引发的查询取消请求路由错误问题，该方案引入了 Peering 机制进行解决 [1]。在基于 AWS EC2 c7i.4xlarge 实例进行的基准测试中，单进程配置的 PgBouncer 峰值吞吐量约为每秒 87,000 笔交易（transactions/sec），而采用 16 个进程的集群配置后，该数值提升至约每秒 336,000 笔交易 [1]。这一改进使整体吞吐量达到原来的四倍左右 [1]。此外，CPU 利用率数据也显示出显著差异：单进程模式下 CPU 占用率仅为 9% 至 16%，而集群模式下的 CPU 利用率则提升至 52% 至 60%[1]。
ClickHouse Managed Postgres has addressed the performance bottleneck of single-threaded PgBouncer on multi-core CPUs by deploying a cluster of PgBouncer processes that scales proportionally with available cores [1]. This configuration binds all processes to the same port and enables the so_reuseport Linux kernel feature, allowing them to share network connections effectively [1]. To resolve routing errors caused by internal kernel load balancing where query cancellation requests were incorrectly directed, a Peering mechanism was implemented between the processes [1].
Benchmark tests conducted on AWS EC2 c7i.4xlarge instances demonstrated that while a single PgBouncer process achieved a peak throughput of approximately 87k transactions/sec with CPU utilization ranging from 9% to 16%, a cluster of 16 processes reached about 336k transactions/sec [1]. This scaling approach increased overall throughput by roughly four times and improved server resource efficiency, raising CPU utilization in the clustered configuration to between 52% and 60% compared to the single-process setup [1][1].