-
Requests generator (wrk):
- 2048 and 16384 connections tests
- 45s duration per test
- 36 threads/18 cores for the application on NUMA0
- 6 cores assigned to IRQ handling on NUMA0
- Measures req/s, throughput, latency (min, max, mean, 99.9th, 99.999th)
-
Upstream Cluster (Nginx):
- In-memory direct response of 12B size
- 17 cores assigned to the application on NUMA1
- 7 cores assigned to IRQ handling on NUMA1
- 2.6M RPS a hard-limit of the server
- Measures #connections established upstream
-
DUT (Proxy):
- Comparison Envoy, Ng-proxy, and Nginx
- Scaling tested with 1, 2, 4, 8, 16, 24, 32 cores assigned to the proxy
- TCP termination toward single upstream cluster
-
DUT (Tunings):
- IRQ affinity
- 8 cores reserved for the IRQ affinity.
- 24 cores allocated for proxy (per NUMA node).
-
All interfaces are 25Gbps Mellanox and Intel NICs