memony is not recovery andLoadNetwork is too slow with clDnnPlugin

[Hou_y_1]

I met two problem using clDnnPlugin of openvino：

1.  LoadNetwork is too slow; I changle my model with --data_type FP16 ; the last IR model is 90MB; Code as follow:

    Core ie;;
    network = ie->ReadNetwork("model.xml", "model.bin");
    ExecutableNetwork executable_network = ie->LoadNetwork(network, "GPU");

2.  I have wrapped inference_engine to my application;  the memony is not recovery after call my destructor.  do I need to explicitly call destructors of "Core", "CNNNetwork" ... 

I use the issue openvino_2020.2.120 in ubuntu 16.04； I have attached my code.

How can i solve these problems, Looking forward to reply!

[SIRIGIRI_V_Intel]

Hi Hou,

To increase the loading time, you can use cl_cache feature. However for the first time, the load time will be same and then faster. For more information on cl_cache, refer FAQ (OpenCL).

Please answer the below questions with respect to memory recovery issue:

1. Do you see this issue while running on CPU as well?
2. Have you seen this issue in previous versions of OpenVINO?
3. How do you observed that memory is not released?

Regards,

Ram prasad

[Hou_y_1]

Ram prasad (Intel) wrote:

Hi Hou,

To increase the loading time, you can use cl_cache feature. However for the first time, the load time will be same and then faster. For more information on cl_cache, refer FAQ (OpenCL).

Please answer the below questions with respect to memory recovery issue:

1. Do you see this issue while running on CPU as well?
2. Have you seen this issue in previous versions of OpenVINO?
3. How do you observed that memory is not released?

Regards,

Ram prasad

Thanks for your attention.  sugguestion for the first question, I will try, thanks again.

About memory recovery:

3.  I observed memory by executing "top" (MEM) and "free" (available)

1.  I use the same code running on CPU,  after calling my wrap class destructor (note: my application not exited),  the MEM dropped from 4.2% to 1.2%,  but GPU from 5.5% to 5.3%.  The  available mem of  "free" is similar . Meanwhile, If I create and release wrap object circularly, the memory increases continuely. the Code in attachment. 

2.  I do not compare with previous versions of OpenVINO, 

[SIRIGIRI_V_Intel]

Please refer to the object_detection_demo_ssd_async demo and try a similar approach on your application too. 

Regards,

Ram prasad