Re: [lwip-users] Help in getting maximum throughput with LWIP.

[Dany Thiffeault]

ok, thanks to David for the nice explanation. I was a little lost!

Chris and Simon, thanks too. I will try the 1516 size right away.

I found in the AVR32 datasheet this line:

• Automatic Pad and CRC Generation on Transmitted Frames

I guess it's what I want, but it doesn't seem intuitive to enable/disable this feature, so I'm wondering if it's what I'm looking for. I will probably try to implement my own memcpy and checksum functions in assembler, but I must say I'm a little scared. The last time I used assembler is like 10 years ago in a simple school project hehe.

The other thing I may try in the meantime is to use the LWIP_CHECKSUM_ON_COPY feature, but since I'm using 1.3.2, I will have to do an update and well... Is there a real improvement in using this option?

The other thing I tried to enable was the MEM_USE_POOLS, but I can't build my project when set to 1. I get errors in the lwip framework and can't find the reason so far. I must miss something.

Thanks!

On Wed, Mar 31, 2010 at 2:15 AM, Simon Goldschmidt <address@hidden> wrote:

Dany Thiffeault wrote:
> oh and Simon, another question. I haven't seen anything about this in your
> wiki, but are PBUF_POOL_SIZE and PBUF_POOL_BUFSIZE important too?

Oh, then it seems like I forgot those two: PBUF_POOL_SIZE is the number of PBUF_POOL pbufs, each with a size of PBUF_POOL_BUFSIZE bytes.

> I'm trying to put everything as high as possible, but I have linking
> problems because I need to much memory. Could you give me some suggestions
> on the parameters to choose if I want to send packets of 8k size, or more,

Maximum packet size over ethernet will be 1514 bytes, so a PBUF_POOL_BUFSIZE of 1516 will be enough. However, if you are low on memory, setting this smaller (e.g. 512 or 256) help to save memory: smaller packets waste less RAM and larger packets will use a chain of PBUF_POOLs (e.g. eith 512, a full-sized packet of 1514 bytes will use 3 pbufs) (- this is what Chris just wrote about). Although the chaining works quite well, there's still a small performance gain of having the packet in one piece, so maximum throughput will be achieved with 1516.

However, there are other places in the stack that have more impact on throughput, mainly checksum generation/checking and memcpy. These should  be optimized for your architecture.

Simon
--
GMX DSL: Internet, Telefon und Entertainment für nur 19,99 EUR/mtl.!
http://portal.gmx.net/de/go/dsl02

_______________________________________________
lwip-users mailing list
address@hidden
http://lists.nongnu.org/mailman/listinfo/lwip-users