[Important] MKLDNN version of the code must have all space of memory

[mozga-intel]

Problem with the MKLDNN performance after this change : commit

Description and solution
The mkldnn performance of the newest Paddle is a worse than the code before the changes: nearly 50%, The mkldnn version of code need to have all space of the memory to work correctly. During the tests with htop the loss of power of the threads is significant and we can not keep the fluency between the threads.

The code shows the solution to this problem. Summary: this is the fall back to the previous version of the code + MKLDNN flag.

[luotao1]

@mozga-intel I couldn't open

[mozga-intel]

@luotao1. I am sorry. The link is up-to-date

[tensor-tang]

Thanks @mozga-intel.

But I am afraid you can not change like this, because initial_cpu_memory_in_mb should also be used when WITH_MKLDNN.

The performance drop is only caused by the changes of memory size. You can try to get the necessary memory size for MKLDNN, like 1G not 500M. Then change the default size with ifdef PADDLE_WITH_MKLDNN.

Because online service also need the memory as a target, so we should keep as small memory as possible, The original default flag would take about 3% of system memory, which would be too large sometimes, like 6G,

That is the reason we add this flag.

[tensor-tang]

Please try to keep this flag available no matter with or without MKLDNN.

The recommend way: change the default size when use MKLDNN. At least, we can know MKLDNN need how much memory to get the best performance.

[tensor-tang]

@mozga-intel
I can refine it later, please wait a moment.

[mozga-intel]

@tensor-tang Thank you for your quickly answer.

I stay in touch with you.