(Measured on Mar 28, 2019)
Used the same settings unless otherwise noted.
- 3 replicas: VMs in the same datacenter (8 cores, 16GB RAM).
- Quorum write: commit a log when 2 out of 3 nodes agree on it.
- Average network RTT: ~180 us
- No client batching.
- A single request always corresponds to a single Raft log.
- Payload size: 256 bytes.
- Benchmark program
- Throughput of a single client thread.
- Send next request only after getting response of previous request.
- Replication throughput is saturated at roughly 5,000 ops/sec.
| Percentile | Latency (us) |
|---|---|
| 50%-tile | 187 |
| 99%-tile | 478 |
| 99.9%-tile | 649 |
| 99.99%-tile | 3746 |
| 99.999%-tile | 8192 |
- Multiple threads are issuing replication requests on the same Raft group.
- It will be saturated by either network bandwidth or CPU capacity.
- Single thread throughput, varying the payload size from 128 bytes to 16 MB.
- Overall throughput is saturated at 320 MB/s.
