Too high CPU usage limit GPU performance

[Yabo_W_]

Hi all,

In my transcoder application, video decode and preprocess(rescale, deinterlace) base on ffmpeg/libav(use cpu), and video encode base on Intel QSV(use GPU).

When the CPU is full(nearly 100%), the Inter QSV encode performance is slowdown, but the GPU usage is not very high, below is the metrics_monitor output:

I want to know whether full CPU usage can slowdown GPU performance .

first, I use sample_encode_drm do some test:

@localhost samples]$ time sudo ./__cmake/intel64.make.release/__bin/release/sample_encode_drm h264 -i ~/input_1080P.yuv -o ~/output.h264 -w 1920 -h 1080 
libva info: VA-API version 0.35.0
libva info: va_getDriverName() returns 0
libva info: Trying to open /usr/lib64/dri/i965_drv_video.so
libva info: Found init function __vaDriverInit_0_32
libva info: va_openDriver() returns 0
Encoding Sample Version 0.0.000.0000

... ...

Processing started
Frame number: 1000
Processing finished

real	0m3.217s
user	0m0.982s
sys	0m0.690s

real time is 0m3.217s

second, I use ffmpeg to encode a 1920x1080 video with best quality to make the CPU full(nearly 100%), and use sample_encode_drm to test encode again:

@localhost samples]$ time sudo ./__cmake/intel64.make.release/__bin/release/sample_encode_drm h264 -i ~/input_1080P.yuv -o ~/output.h264 -w 1920 -h 1080 
libva info: VA-API version 0.35.0
libva info: va_getDriverName() returns 0
libva info: Trying to open /usr/lib64/dri/i965_drv_video.so
libva info: Found init function __vaDriverInit_0_32
libva info: va_openDriver() returns 0
Encoding Sample Version 0.0.000.0000

... ...

Processing started
Frame number: 1000
Processing finished

real	0m3.509s
user	0m2.012s
sys	0m1.221s

real time is 0m3.509s

Of course, CPU is full, the I/O performance maybe slower.

Sorry for my bad English.

[Surbhi_M_Intel · Accepted solution]

Hi Yabo, 

I don't think occupying CPU completely should reduce Graphics performance, they work independently of each other workload. One reason why your test is not reflecting same because in sample_encode_drm, it is reading an input from a file and writing the o/p which involves file I/O and done on core, also it is not optimized enough which could reduce the performance of your encode test if CPU  is fully occupied. You can test without counting the read and write function in your test(only GPU is being used) and check the test time or fps in both scenarios when your CPU is idle and completely loaded. 

Thanks,
Surbhi

[Yabo_W_]

Hi Surbhi,

I had do some test as your suggestion. I add some code to sample_encode/src/pipeline_encode.cpp:

1192 
1193 + static long getCurrentTime()
1194 + {
1195 +    struct timeval tv;
1196 +    gettimeofday(&tv,NULL);
1197 +    return tv.tv_sec * 1000000 + tv.tv_usec;
1198 + }

1314         for (;;)
1315         {
1316 +            long start_ms = getCurrentTime();
1317              // at this point surface for encoder contains either a frame from file or a frame processed by vpp
1318              sts = m_pmfxENC->EncodeFrameAsync(NULL, &m_pEncSurfaces[nEncSurfIdx], &pCurrentTask->mfxBS, &pCurrentTask->EncSyncP);
1319 +            fprintf(stderr, "EncodeFrameAsync() take %lf\n", (getCurrentTime() - start_ms) / (double)1000000);

If the CPU is idle, sample_encode_drm output:

EncodeFrameAsync() take 0.000006
EncodeFrameAsync() take 0.000006
EncodeFrameAsync() take 0.000006
EncodeFrameAsync() take 0.000006
EncodeFrameAsync() take 0.000005
EncodeFrameAsync() take 0.000003
EncodeFrameAsync() take 0.000003
EncodeFrameAsync() take 0.000003
EncodeFrameAsync() take 0.000003
EncodeFrameAsync() take 0.000003
EncodeFrameAsync() take 0.000003
EncodeFrameAsync() take 0.000006
EncodeFrameAsync() take 0.000005
EncodeFrameAsync() take 0.000006
EncodeFrameAsync() take 0.000006
EncodeFrameAsync() take 0.000005
EncodeFrameAsync() take 0.000005
EncodeFrameAsync() take 0.000004
EncodeFrameAsync() take 0.000003
EncodeFrameAsync() take 0.000003
EncodeFrameAsync() take 0.000003

If the CPU is completely loaded,sample_encode_drm output:

EncodeFrameAsync() take 0.000008
EncodeFrameAsync() take 0.000008
EncodeFrameAsync() take 0.000007
EncodeFrameAsync() take 0.000020
EncodeFrameAsync() take 0.000018
EncodeFrameAsync() take 0.000009
EncodeFrameAsync() take 0.000025
EncodeFrameAsync() take 0.000016
EncodeFrameAsync() take 0.000009
EncodeFrameAsync() take 0.000021
EncodeFrameAsync() take 0.000007
EncodeFrameAsync() take 0.000009
EncodeFrameAsync() take 0.000032
EncodeFrameAsync() take 0.000017
EncodeFrameAsync() take 0.000009
EncodeFrameAsync() take 0.000018
EncodeFrameAsync() take 0.000018
EncodeFrameAsync() take 0.000010
EncodeFrameAsync() take 0.000023
EncodeFrameAsync() take 0.000009
EncodeFrameAsync() take 0.000009

completely loaded CPU take more time than idle CPU.