Add design documentation for profiling tool. #6700

qingqing01 · 2017-12-18T06:40:33Z

For a better review, you could see it here.

reyoung · 2017-12-21T02:55:59Z

doc/design/profiler.md

+  kPushRange,
+  kPopRange};
+```
+- kMark: only a mark.


I do not understand what is only a mark.

Add more comments.

reyoung · 2017-12-21T02:56:38Z

doc/design/profiler.md

+In above work flow, a pair of events are needed before and aftern the piece of code to collect time. So the event has a flag to mark it is starting event or ending event. There three kinds of event:
+
+```c++
+enum EventKind { kMark,


Better Indent

reyoung · 2017-12-21T02:56:57Z

doc/design/profiler.md

+  kCPU,
+  kCUDA
+};
+ProfilerState kState;


Done. And also update in the code.

kuke

A great design doc. Thanks!

kuke · 2017-12-21T08:56:21Z

doc/design/profiler.md

@@ -0,0 +1,95 @@
+## Introduction
+
+There are many performance analysis tools for [different programming languages and different software framework](https://en.wikipedia.org/wiki/List_of_performance_analysis_tools). For most popular deep learning framework, they used several programming languages and adapt to heterogeneous platforms. Similar to most of the deep learning framework, the PaddlePaddle also used C++, CUDA and Python as the basic programming languages to adapt to run on CPU and GPU device.  The [`nvprof` tools](http://docs.nvidia.com/cuda/profiler-users-guide/index.html#nvprof-overview) is usually used to analyse the CUDA program.  We have [a document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/optimization/cpu_profiling.md) to profile CPU and Python program by [yep](https://pypi.python.org/pypi/yep) and [Google's perftools](https://github.com/google/pprof) to profile the only CPU and Python program. But for [PaddlePaddle fluid](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/fluid.md), the operator is the basic computing unit. The developers usually wants to collect the time of each operator and locate bottlenecks.  The `nvprof` usually collect the timeline of CUDA-related activities on both CPU and GPU, including kernel execution, memory transfers, memory set and CUDA API calls and events or metrics for CUDA kernels. And the `yep` and `Google's perftools` can't collect the timeline for CUDA program. All these tools can't collect time in the operator level. So we design this profiling tools.


framework -> frameworks when the plural form is needed. The same below.
they used -> they use
the PaddlePaddle -> PaddlePaddle
CPU and GPU device -> CPU and GPU devices
the only -> only the
profiling tools -> profiling tool

Done. Thanks!

kuke · 2017-12-21T09:18:38Z

doc/design/profiler.md

+
+## Architecture
+
+The work flow for most task is as follows. Each operator will run many times in the all iterations. So the profiler must collect the total time of each operator during the iteration. For more, sometimes, the developers want to collect more detailed time span inside the operator or record time span for elsewhere, this requires that the profiler must support to record the nested time span. And in order to speed training, all the deep learning framework supports parallel computing, including multi-threads on CPU and multi-GPUs. So the profiler must enable to collect the timeline for each thread. In addition, the profiler also occupies certain resources. It must can be easily to enable or disable by the developers. At last, the profiler should show a human-readable report.  


speed -> speedup
multi-threads, multi-GPUs -> multiple threads, multiple GPUs. They should be nouns here.
enable -> be able
It must can be easily to enable or disable by the developers -> It must be easily to be enabled or disabled by the developers
show -> present

kuke · 2017-12-21T09:20:19Z

doc/design/profiler.md

+
+```python
+for i in xrange(M):  # M is  the iteration number
+  for op in operator_lists: # The `operator_lists` is the all operators in the network graph.


is the all operators -> contains all the operators

kuke · 2017-12-21T09:21:51Z

doc/design/profiler.md

+    op.run();
+```
+
+In a summary, the proflier should have follow features:


In a summary -> In summary
follow features -> the following features

kuke · 2017-12-21T09:23:03Z

doc/design/profiler.md

+
+In a summary, the proflier should have follow features:
+
+- record time span in loop.


record -> records
support -> supports

kuke · 2017-12-21T09:33:07Z

doc/design/profiler.md

+
+But how to record the time for the mixed C++ and CUDA program?  There many C++ interfaces to get the current calendar time in host program. But for GPU, the CUDA kernels may be executed concurrently if they are in different streams (http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#streams) and the CUDA kernels is asynchronous with the host program if there is no the synchronous aftern the CUDA kernels. The CUDA provides [event](http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#events) to monitor the device's perform accurate timing. Inspired by PyTorch and CUDA event, we also design and apply the events to record the timeline.  Then summary and show statistics based on these events.  
+
+The overall flow is shown as following figure.


as following figure -> as the following figure

kuke · 2017-12-21T09:36:29Z

doc/design/profiler.md

+
+### Event
+
+In above work flow, a pair of events are needed before and aftern the piece of code to collect time. So the event has a flag to mark it is starting event or ending event. There three kinds of event:


aftern -> after
to collect time -> to be analyzed
mark it is starting event or ending event -> mark whether it is a starting event or an ending event
There -> There are

Done. Thanks!

kuke · 2017-12-21T09:38:26Z

doc/design/profiler.md

+```
+- kMark: only a mark.
+- kPushRange: mark the starting event for time range. 
+- kPopRange: mark the ending event for the time range.


the time range -> time range for consistency

kuke · 2017-12-21T09:39:49Z

doc/design/profiler.md

+- kPushRange: mark the starting event for time range. 
+- kPopRange: mark the ending event for the time range.
+
+For the CPU code, the events only need to record the current time. For the CUDA code, the [event management functions of CUDA](http://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__EVENT.html#group__CUDART__EVENT) are used.  For many pieces of code, a event lists are used to record each piece.


a event lists are used to record each piece-> an event list is used to record their timeline

kuke · 2017-12-21T09:44:51Z

doc/design/profiler.md

+};
+```
+
+As mentioned above, there is no need to record the timeline when disabling the profiler. So there is a global state to enable or distable the profiler. 


distable -> disable

kuke

LGTM

Add profiler design documentation.

7c63eaa

qingqing01 requested review from chengduoZH, kuke and reyoung December 18, 2017 06:40

qingqing01 mentioned this pull request Dec 18, 2017

[Profiling tool]: Add the event to record the timeline. #6701

Merged

qingqing01 requested a review from dzhwinter December 21, 2017 02:28

reyoung reviewed Dec 21, 2017

View reviewed changes

kuke reviewed Dec 21, 2017

View reviewed changes

Fix the typo error and add more comments.

2a5c6a4

qingqing01 force-pushed the profiler_design branch from 7a1dc86 to 2a5c6a4 Compare January 2, 2018 02:08

kuke approved these changes Jan 2, 2018

View reviewed changes

qingqing01 merged commit 3ebfd7f into PaddlePaddle:develop Jan 3, 2018

qingqing01 deleted the profiler_design branch November 14, 2019 05:28

		@@ -0,0 +1,95 @@
		## Introduction

		There are many performance analysis tools for [different programming languages and different software framework](https://en.wikipedia.org/wiki/List_of_performance_analysis_tools). For most popular deep learning framework, they used several programming languages and adapt to heterogeneous platforms. Similar to most of the deep learning framework, the PaddlePaddle also used C++, CUDA and Python as the basic programming languages to adapt to run on CPU and GPU device. The [`nvprof` tools](http://docs.nvidia.com/cuda/profiler-users-guide/index.html#nvprof-overview) is usually used to analyse the CUDA program. We have [a document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/optimization/cpu_profiling.md) to profile CPU and Python program by [yep](https://pypi.python.org/pypi/yep) and [Google's perftools](https://github.com/google/pprof) to profile the only CPU and Python program. But for [PaddlePaddle fluid](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/fluid.md), the operator is the basic computing unit. The developers usually wants to collect the time of each operator and locate bottlenecks. The `nvprof` usually collect the timeline of CUDA-related activities on both CPU and GPU, including kernel execution, memory transfers, memory set and CUDA API calls and events or metrics for CUDA kernels. And the `yep` and `Google's perftools` can't collect the timeline for CUDA program. All these tools can't collect time in the operator level. So we design this profiling tools.


		## Architecture

		The work flow for most task is as follows. Each operator will run many times in the all iterations. So the profiler must collect the total time of each operator during the iteration. For more, sometimes, the developers want to collect more detailed time span inside the operator or record time span for elsewhere, this requires that the profiler must support to record the nested time span. And in order to speed training, all the deep learning framework supports parallel computing, including multi-threads on CPU and multi-GPUs. So the profiler must enable to collect the timeline for each thread. In addition, the profiler also occupies certain resources. It must can be easily to enable or disable by the developers. At last, the profiler should show a human-readable report.


		In a summary, the proflier should have follow features:

		- record time span in loop.


		But how to record the time for the mixed C++ and CUDA program? There many C++ interfaces to get the current calendar time in host program. But for GPU, the CUDA kernels may be executed concurrently if they are in different streams (http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#streams) and the CUDA kernels is asynchronous with the host program if there is no the synchronous aftern the CUDA kernels. The CUDA provides [event](http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#events) to monitor the device's perform accurate timing. Inspired by PyTorch and CUDA event, we also design and apply the events to record the timeline. Then summary and show statistics based on these events.

		The overall flow is shown as following figure.


		### Event

		In above work flow, a pair of events are needed before and aftern the piece of code to collect time. So the event has a flag to mark it is starting event or ending event. There three kinds of event:

Add design documentation for profiling tool. #6700

Add design documentation for profiling tool. #6700

Uh oh!

Conversation

qingqing01 commented Dec 18, 2017

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kuke left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kuke left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants