RE: FEC performance degradation with certain packet sizes

["fugang.duan@freescale.com" <fugang.duan@freescale.com>]

From: Hector Palacios <hector.palacios@digi.com>
Sent: Friday, December 20, 2013 11:02 PM

>To: Duan Fugang-B38611; Marek Vasut; netdev@vger.kernel.org
>Cc: Estevam Fabio-R49496; shawn.guo@linaro.org; l.stach@pengutronix.de; Li
>Frank-B20596; bhutchings@solarflare.com; davem@davemloft.net
>Subject: Re: FEC performance degradation with certain packet sizes
>
>Dear Andy,
>
>On 12/20/2013 04:35 AM, fugang.duan@freescale.com wrote:
>> [...]
>>
>> I can reproduce the issue on imx6q/dl platform with freescale internal kernel
>tree.
>>
>> This issue must be related to cpufreq, when set scaling_governor to
>performance:
>> echo performance >
>> /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
>>
>> And then do NPtcp test, the result as below:
>>
>>   24:      99 bytes      5 times -->      9.89 Mbps in      76.40 usec
>>   25:     125 bytes      5 times -->     12.10 Mbps in      78.80 usec
>>   26:     128 bytes      5 times -->     12.27 Mbps in      79.60 usec
>>   27:     131 bytes      5 times -->     12.80 Mbps in      78.10 usec
>>   28:     189 bytes      5 times -->     18.00 Mbps in      80.10 usec
>>   29:     192 bytes      5 times -->     18.31 Mbps in      80.00 usec
>>   30:     195 bytes      5 times -->     18.41 Mbps in      80.80 usec
>>   31:     253 bytes      5 times -->     23.34 Mbps in      82.70 usec
>>   32:     256 bytes      5 times -->     23.91 Mbps in      81.70 usec
>>   33:     259 bytes      5 times -->     24.19 Mbps in      81.70 usec
>>   34:     381 bytes      5 times -->     33.18 Mbps in      87.60 usec
>>   35:     384 bytes      5 times -->     33.87 Mbps in      86.50 usec
>>   36:     387 bytes      5 times -->     34.41 Mbps in      85.80 usec
>>   37:     509 bytes      5 times -->     42.72 Mbps in      90.90 usec
>>   38:     512 bytes      5 times -->     42.60 Mbps in      91.70 usec
>>   39:     515 bytes      5 times -->     42.80 Mbps in      91.80 usec
>>   40:     765 bytes      5 times -->     56.45 Mbps in     103.40 usec
>>   41:     768 bytes      5 times -->     57.11 Mbps in     102.60 usec
>>   42:     771 bytes      5 times -->     57.22 Mbps in     102.80 usec
>>   43:    1021 bytes      5 times -->     70.69 Mbps in     110.20 usec
>>   44:    1024 bytes      5 times -->     70.70 Mbps in     110.50 usec
>>   45:    1027 bytes      5 times -->     69.59 Mbps in     112.60 usec
>>   46:    1533 bytes      5 times -->     73.56 Mbps in     159.00 usec
>>   47:    1536 bytes      5 times -->     72.92 Mbps in     160.70 usec
>>   48:    1539 bytes      5 times -->     73.80 Mbps in     159.10 usec
>>   49:    2045 bytes      5 times -->     93.59 Mbps in     166.70 usec
>>   50:    2048 bytes      5 times -->     94.07 Mbps in     166.10 usec
>>   51:    2051 bytes      5 times -->     92.92 Mbps in     168.40 usec
>>   52:    3069 bytes      5 times -->    123.43 Mbps in     189.70 usec
>>   53:    3072 bytes      5 times -->    123.68 Mbps in     189.50 usec
>
>You are right. Unfortunately, this does not work on i.MX28 (at least for me).
>Couldn't it be that the cpufreq is masking the problem on the i.MX6?
>
>Best regards,
>--
>Hector Palacios
>

I reproduce the issue on imx28 evk platform, imx28 has no specific cpufreq driver. 
In kernel 3.13, the ethernet driver is almost the same for imx28 and imx6 since they use the
Same enet IP, but imx6 enet IP have some evolution.

Now I don't know the cause. When I am free, I will dig out it.

Thanks,
Andy