Benchmark for different kind of workload, and different setups.
Machines are located close to each other, in the same data center.
DPDK server serves a http html page, and tokio test client try to make as much traffic as possible. See:
Driver:
mlx5_core: Ethernet controller: Mellanox Technologies MT27800 Family [ConnectX-5 Virtual Function]
Azure Bench Result:
- Standard_D2s_v5: 2 cpu, 2 queue pair.
- Standard_D4s_v5: 4 cpu, 4 queue pair.
- Standard_D8s_v5: 8 cpu, 8 queue pair.
Note:
- In some vm queue is less than cpu, so the machine is not fully utilized by the current bench server, which creates 1 thread per queue and process http request on that thread. But the above vm sizes has 1 queue per cpu.
- The bench-test tokio client runs on
Standard_D8s_v5VM. Itself has perf bottle neck, and manifests in theStandard_D8s_v5server tests. (TODO: use DPDK as the test client.)