diff --git a/cube/doc/performance.md b/cube/doc/performance.md
index d250dae0fc79635f705144d0306225a9842c0a53..5fa297772c5fb3a0668e4a2a2f720b6f0a079ff7 100644
--- a/cube/doc/performance.md
+++ b/cube/doc/performance.md
@@ -10,13 +10,15 @@ Intel(R) Xeon(R) CPU E5-2650 v4 @ 2.20GHz
| qps | 10w | 50w | 100w |
-| --- | --- | --- | --- |
+|---|---|---|---|
|kps|10w|50w|100w|
|cpu(top)| 6.5% | 38.3% | 71.4% |
|client端延迟| avg 196 us
50% 160 us
70% 188 us
90% 292 us
95% 419 us
97% 547 us
99% 835 us
99.9% 1556 us
99.99% 1779 us| avg 563 us
50% 342 us
70% 502 us
90% 1063 us
95% 1703 us
97% 2399 us
99% 4036 us
99.9% 7195 us
99.99% 7340 us| avg 4234 us
50% 3120 us
70% 5459 us
90% 10657 us
95% 14074 us
97% 16215 us
99% 19434 us
99.9% 29398 us
99.99% 33921 us|
## 高kps场景
生产环境下，预估服务更多的会以batch形式访问cube server，这类情况kps相对qps能更准确的判断服务性能。我们以单次100key为例，给出不同kps下cube server的相关性能指标。
+
+
| qps | 2w | 10w | 20w |
|---|---|---|---|
|kps|200w|1000w|2000w|
@@ -29,6 +31,8 @@ Intel(R) Xeon(R) CPU E5-2650 v4 @ 2.20GHz
测试条件:
cube分片数:10
client机器ping下游server机器约0.06ms
+
+
| batch size | 100 | 500 | 1000 |
|---|---|---|---|
|qps|100|100|100|