From 9e3741dcce387d9276cf749d5f9c20b4640dea68 Mon Sep 17 00:00:00 2001
From: lfu-ps <83976250+lfu-ps@users.noreply.github.com>
Date: Mon, 1 Aug 2022 20:35:00 -0700
Subject: [PATCH] Create Arion_Agent_Benchmark.MD (#2)

---
 benchmark/Arion_Agent_Benchmark.MD | 62 ++++++++++++++++++++++++++++++
 1 file changed, 62 insertions(+)
 create mode 100644 benchmark/Arion_Agent_Benchmark.MD

diff --git a/benchmark/Arion_Agent_Benchmark.MD b/benchmark/Arion_Agent_Benchmark.MD
new file mode 100644
index 0000000..0abd060
--- /dev/null
+++ b/benchmark/Arion_Agent_Benchmark.MD
@@ -0,0 +1,62 @@
+# ArionAgent performance
+## Overview
+
+This is the performance test of Arion Agent working with ebpf/XDP as downstream programming module and ArionMaster grpc streaming server as upstream metadata source as an entire system.
+
+
+## Test environment
+
+This test is between 2 machines of the same lab, they don't belong to the same IP range (means not located on the same rack but same data center):
+    
+    1. Arion master server
+    
+    2. Arion agent (launched on the same machine of Arion Wing which is XDP as gateway network functionality)
+
+
+## Test workflow and scenario #1 - watch from Arion Master only latency (ArionAgent as c++ grpc client)
+
+Latency = ArionAgent finishes receiving N number of neighbors - start to receive Grpc neighbors time (right after watch call to ArionMaster Grpc server finished)
+
+Watch performance from ArionMaster:
+* Watch 5 neighbors performance: 31 us
+* Watch 100k neighbors performance: 379,270 us = ~380 ms
+
+
+## Test workflow and scenario #2 - E2E ebpf programming latency
+
+Latency = Finish ebpf map programming time - start to receive Grpc neighbors time (right after watch call to ArionMaster Grpc server finished)
+
+E2E programming (watch + programming) latency:
+* 1 neighbor
+    - overall latency 15us
+    - per neighbor latency 15us
+* 5 neighbors
+    - overall latency 61us
+    - per neighbor latency 12.2us
+* 100 neighbors
+    - overall latency 1004us = 1ms
+    - per neighbor latency 10us
+* 500 neighbors
+    - overall latency 4803us = 4.8ms
+    - per neighbor latency 9.6us
+* 1k neighbors
+    - overall latency 9.5ms
+    - per neighbor latency 9.5us 
+* 5k neighbors
+    - overall latency 49.1ms
+    - per neighbor latency 9.8us
+* 10k neighbors
+    - overall latency 93.6ms
+    - per neighbor latency 9.36us
+* 50k neighbors
+    - overall latency 354.1ms
+    - per neighbor latency 7.08us
+* 100k neighbors
+    - overall latency 601.4ms (and if we compare with the watch only 100k neighbors latency of 380ms, we know the overhead of 100k ebpf rule programming is 220ms)
+    - per neighbor latency 6.01us
+* 500k neighbors
+    - overall latency 2924.6ms = 2.9 seconds
+    - per neighbor latency 5.85us
+* 1 million neighbors
+    - overall latency 5,893 ms = 5.9 seconds
+    - per neighbor latency 5.89us