From 5dbd537048f153863d1cac0fbb433bb62fe561b4 Mon Sep 17 00:00:00 2001
From: Yancey <yanxu05@baidu.com>
Date: Fri, 12 Jan 2018 11:25:53 +0800
Subject: [PATCH] Fluid distributed training benchmark (#7410)

* add cluster training bencharmk design

* update by comment

* update by comment
---
 benchmark/cluster/README.md | 78 +++++++++++++++++++++++++++++++++++++
 1 file changed, 78 insertions(+)
 create mode 100644 benchmark/cluster/README.md

diff --git a/benchmark/cluster/README.md b/benchmark/cluster/README.md
new file mode 100644
index 000000000..b619613ea
--- /dev/null
+++ b/benchmark/cluster/README.md
@@ -0,0 +1,78 @@
+# Cluster Training Benchmark
+
+## Setup
+
+- Platform
+  - Kubernetes: v1.6.2
+  - Linux Kernel: v3.10.0
+
+- Resource
+  - CPU: 10 Cores per Pod
+  - Memory: 5GB per Pod
+
+- Docker Image
+
+  We use different base Docker Image to run the benchmark on Kubernetes:
+  - PaddlePaddle v2: paddlepaddle/paddle:0.11.0
+  - PaddlePaddle Fluid: paddlepaddle/paddle:[commit-id]
+  - TensorFlow: tensorflow/tensorflow:1.5.0-rc0
+
+- Model
+  vgg16 is used in this benchmark.
+
+## Cases
+
+- Variable
+  - Batch Size of training data.
+  - PServer count of the training job.
+  - The number of trainers.
+
+- Invariant
+  - The resource of trainer/pserver Pod.
+
+### Measure the Performance for Different Batch Size
+
+- PServer Count: 40
+- Trainer Count: 100
+- Metrics: mini-batch / sec
+
+| Batch Size | 32 | 64 | 128 | 256 |
+| -- | -- | -- | -- | -- |
+| PaddlePaddle Fluid | - | - | - | - |
+| PaddlePaddle v2 | - | - | - | - |
+| TensorFlow | - | - | - | - |
+
+### Measure the Performance for Different PServer Count
+
+- Trainer Count: 100
+- Batch Size: 64
+- Metrics: mini-batch / sec
+
+| PServer Count | 10 | 20 | 40 | 60 |
+| -- | -- | -- | -- | -- |
+| PaddlePaddle Fluid | - | - | - | - |
+| PaddlePaddle v2 | - | - | - | - |
+| TensorFlow | - | - | - | - |
+
+### Measure Parallel Efficiency By Increasing Trainer Count
+
+- PServer Count: 20
+- Batch Size: 64
+- Metrics:
+
+$S = \div(T1, TN)$
+
+which S is the ratio of T1 over TN, training time of 1 and N trainers.
+The parallel efficiency is:
+
+$E = \div(S, N)$
+
+| Trainer Counter | 1 | 10 | 20 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 |
+| -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- |
+| PaddlePaddle Fluid | - | - | - | - | - | - | - | - | - | - | - |
+| PaddlePaddle v2 | - | - | - | - | - | - | - | - | - | - | - | - |
+| TensorFlow | - | - | - | - | - | - | - | - | - | - | - | - | - |
+
+## Reproduce the benchmark
+
+TODO
-- 
GitLab