Re: [lwip-users] How to optimize raw UDP performance

[Max Bobrov]

I am mainly concerned about transmitting at at least 150Mbits/sec.

The speeds i'm refering to are provided by xilinx in the xapp1026.pdf
document they provide for using lwip on a xilinx board. They claim tcp
in raw mode performs at 140Mbits/sec. I assume UDP is faster, but i'm
no where near the 140 right now.

On Wed, Sep 23, 2009 at 5:03 PM, Bill Auerbach
<address@hidden> wrote:
> Which direction?  Improving sending and improving receiving are 2 different
> problems.  They overlap some, but the focus is different.
>
> Something's not right if offloading was used and it didn't help - it should
> double to quadruple the speed.  Try turning off CHECKSUM_GEN_UDP and
> CHECKSUM_CHECK_UDP.  Offloading or doing the checksum in hardware is one
> critical item.  As is optimizing memory and memory copies (using a DMA
> memcpy over the standard C memcpy is huge).  Using zero-copy drivers for
> sending and/or receiving is huge (because you're not copying).
>
> I spent 2+ months optimizing lwIP - at first for TCP, which although I
> doubled, wasn't nearly fast enough.  We switched to UDP and proved the speed
> was adequate to move forward.  Then I wrote a reliable UDP protocol (not
> hard to do, but had to be done since we can't lose data).  Some of my
> optimizations were submitted but some may be target dependent (i.e. little
> endian improvements or specific to my variant of GCC).  I've pushed the past
> 2 years for better performance and I realize resources are limited to
> support it as well as it's not a top priority for the goals of lwIP (Light
> Weight being the key words here).  For most users I suspect what we have is
> fast enough - plus it's a very good and reliable implementation of TCP/IP.
>
> Documented speed has a lot to do with the platform.  On a high speed PowerPC
> with assembly internet checksum and DMA memory copies, lwIP met our needs
> without any other changes or optimizations except a low level UDP/ICMP/ARP
> packet filter (in low_level_input) while in our "high bandwidth" mode.  We
> didn't have to do this - I just didn't want a UDP or ICMP burst to disrupt
> system timing.  On a lower speed FPGA project, a lot of effort was put into
> lwIP and supporting code and the FPGA to meet our goals.
>
> Bill
>
>
>
>>-----Original Message-----
>>From: address@hidden
>>[mailto:address@hidden On
>>Behalf Of bobbyb
>>Sent: Wednesday, September 23, 2009 4:02 PM
>>To: address@hidden
>>Subject: [lwip-users] How to optimize raw UDP performance
>>
>>
>>What are some good techniques for improving UDP performance in raw mode.
>>Right now i'm getting approximately 82Mbits/sec. I would like to
>>increase
>>that to 150+ Mbits/sec which should be possible according to the
>>documentation. I've tried increasing various buffer/memory sizes with
>>very
>>small effects. I also setup checksum offloading which surprisingly had
>>no
>>effect on performance.
>>
>>Does anyone know any other techniques worth trying? Why is the
>>documented
>>speed so much higher than what I'm getting?
>>
>>Thanks
>>--
>>View this message in context: http://www.nabble.com/How-to-optimize-raw-
>>UDP-performance-tp25531280p25531280.html
>>Sent from the lwip-users mailing list archive at Nabble.com.
>>
>>
>>
>>_______________________________________________
>>lwip-users mailing list
>>address@hidden
>>http://lists.nongnu.org/mailman/listinfo/lwip-users
>
>
>
> _______________________________________________
> lwip-users mailing list
> address@hidden
> http://lists.nongnu.org/mailman/listinfo/lwip-users
>