Add paddle/memory/README.md #2552

wangkuiyi · 2017-06-21T20:20:57Z

No description provided.

gangliao · 2017-06-22T00:40:57Z

paddle/README.md

+In `paddle/memory/memory.h` we have:
+
+```cpp
+template <typeanme Place> void* Alloc(Place, size_t);


typeame -> typename

Since Place is a variant, why not directly use the typeid to static analyze place type during compilation?

void* Alloc(Place, size_t); void* Alloc(Place place, size_t size) { if (place.type() == typeid(GPUPlace)) { .... } else {...} }

Because we want the mismatched-type error get reported as early as possible. Err at compile time is earlier than runtime.

I got it. Thanks.

But variant also supports throw the mismatched-type error during compile time.

Actually, I think if we regulate the way we use GPUPlace and CPUPlace, like here -- two and only two specializations of Alloc<Place>, we might not need typedef boost::variant<GPUPlace, CPUPlace> Place at all. And this might help us remove the dependency to boost.

It's great to think about this. The hard part to remove boost::variant is Dim.

Yeah. Agree. And thanks for reminding of boost::variant. Let's start by minimizing dependencies of each piece of our work.

gangliao · 2017-06-22T00:48:18Z

paddle/README.md

+BuddyAllocator* GetCPUBuddyAllocator() {
+  static BuddyAllocator* a = NULL;
+  if (a == NULL) {
+    a = new BuddyAllocator(new CPUAllocator /*backup allocator*/, ...);


All member functions in CPUAllocator are static functions, I thought it's better if CPUAllocator /CPUAllocator could be a template type of BuddyAllocator

I don't think GPUAllocator::Alloc could be a static method, as it reads GPUAllocator::place_ and writes GPUAllocator::used_.

OK, I guess we do not need the member variables in here.

GPUAllocator::place_ is unnecessary because you got place from memory::Alloc(place ... each time and GPUAllocator::Alloc insides memory::Alloc.

GPUAllocator::used_ is unnecessary. Because each time when you retrieve/query the used memory, it's non-deterministic and likely to change and fluctuate.

I mean, in this design, each GPUAllocator instance corresponds to a GPU device.

Why GPUAllocator::used_ is unnecessary? I think memory profiling is very important for a DL system?

Yeah, It's useful. I mean it's unnecessary to save it in GPUAllocator::used_. because its behavior always to change and fluctuate.

I see. I cannot tell how valuable GPUAllocator::used_ could be. But I'd like to have it, as it enables us to profile the memory usage. We may figure out good usages of this piece of information later.

gangliao · 2017-06-22T00:53:56Z

paddle/README.md

+
+### The Implementation
+
+`GetCPUBuddyAllocator` and `GetGPUBuddyAllocator` are singletions.


Singleton is great in here.

… memory_design

reyoung · 2017-06-22T15:33:22Z

paddle/README.md

+
+```cpp
+template<>
+void Alloc(GPUPlace)(GPUPlace p, size_t size) {


Alloc(GPUPlace) -> Alloc<GPUPlace>

reyoung · 2017-06-22T15:37:17Z

paddle/README.md

+ private:
+  struct Block {
+    size_t size;
+    Blobk* left, right;


Blobk -> Block.

reyoung · 2017-06-22T15:55:51Z

paddle/README.md

+
+The `GPUAllocator` and `CPUAllocator` are calls *system allocators*.  They hold information about the device, including the amount of memory has been allocated.  So that we can call
+
+- `GPUAllocator::Used` and


It seems that System Allocators are used when constructing a BuddyAllocator and they are private data member inside a BuddyAllocator. How can we get GPUAllocator::Used for each device?
Is our code like this?

auto* buddyAllocator = GetGPUBuddyAllocator(0); buddyAllocator->SystemAllocator()->Used();

or

GetGPUAllocator(0)->Used();

Thanks for the reminder. Let me explain more here.

jacquesqiao · 2017-06-23T04:18:09Z

paddle/memory/README.md

@@ -0,0 +1,145 @@
+In my mind, the memory package works like the following:


Should this line be removed?

gangliao

LGTM

Add paddle/memory/README.md

7cb68a8

wangkuiyi requested review from gangliao, JiayiFeng and QiJune June 21, 2017 20:21

Has to auto format networks.py because CI complains about it.

0a92908

gangliao reviewed Jun 22, 2017

View reviewed changes

Yi Wang and others added 2 commits June 22, 2017 01:36

Merge branch 'develop' of https://github.com/paddlepaddle/paddle into…

e732bdd

… memory_design

Revert changes made by misleading errors from Travis CI

d3e2db4

reyoung reviewed Jun 22, 2017

View reviewed changes

wangkuiyi added 2 commits June 22, 2017 10:27

Move README.md from paddle/ to paddle/memory/

8cfa48d

In response to comments from Liao Gang and Yu Yang

c617520

jacquesqiao reviewed Jun 23, 2017

View reviewed changes

Yi Wang and others added 2 commits June 23, 2017 11:42

Remove unnecessary preamble

b55df90

Update design

ab2550c

wangkuiyi mentioned this pull request Jun 26, 2017

Memory CPU allocator #2596

Merged

gangliao approved these changes Jun 26, 2017

View reviewed changes

wangkuiyi merged commit 09eb6be into PaddlePaddle:develop Jun 26, 2017

wangkuiyi deleted the memory_design branch June 26, 2017 20:18

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add paddle/memory/README.md #2552

Add paddle/memory/README.md #2552

wangkuiyi commented Jun 21, 2017

gangliao Jun 22, 2017 •

edited

Loading

wangkuiyi Jun 22, 2017

gangliao Jun 22, 2017

wangkuiyi Jun 22, 2017

gangliao Jun 22, 2017

wangkuiyi Jun 22, 2017 •

edited

Loading

gangliao Jun 22, 2017

wangkuiyi Jun 22, 2017 •

edited

Loading

gangliao Jun 22, 2017

wangkuiyi Jun 22, 2017

wangkuiyi Jun 22, 2017

gangliao Jun 22, 2017 •

edited

Loading

wangkuiyi Jun 22, 2017 •

edited

Loading

gangliao Jun 22, 2017 •

edited

Loading

reyoung Jun 22, 2017

wangkuiyi Jun 22, 2017

reyoung Jun 22, 2017

wangkuiyi Jun 22, 2017

reyoung Jun 22, 2017 •

edited

Loading

wangkuiyi Jun 22, 2017

wangkuiyi Jun 22, 2017

wangkuiyi Jun 23, 2017

jacquesqiao Jun 23, 2017

wangkuiyi Jun 23, 2017

gangliao left a comment


		### The Implementation

		`GetCPUBuddyAllocator` and `GetGPUBuddyAllocator` are singletions.


		The `GPUAllocator` and `CPUAllocator` are calls system allocators. They hold information about the device, including the amount of memory has been allocated. So that we can call

		- `GPUAllocator::Used` and

		@@ -0,0 +1,145 @@
		In my mind, the memory package works like the following:

Add paddle/memory/README.md #2552

Add paddle/memory/README.md #2552

Conversation

wangkuiyi commented Jun 21, 2017

gangliao Jun 22, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wangkuiyi Jun 22, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wangkuiyi Jun 22, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gangliao Jun 22, 2017 • edited Loading

Choose a reason for hiding this comment

wangkuiyi Jun 22, 2017 • edited Loading

Choose a reason for hiding this comment

gangliao Jun 22, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

reyoung Jun 22, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gangliao left a comment

Choose a reason for hiding this comment

gangliao Jun 22, 2017 •

edited

Loading

wangkuiyi Jun 22, 2017 •

edited

Loading

wangkuiyi Jun 22, 2017 •

edited

Loading

gangliao Jun 22, 2017 •

edited

Loading

wangkuiyi Jun 22, 2017 •

edited

Loading

gangliao Jun 22, 2017 •

edited

Loading

reyoung Jun 22, 2017 •

edited

Loading