RE: [lwip-users] LWIP Delay between Consecutive Packets

[Taranowski, Thomas (SWCOE)]

This sounds to me like the driver isn’t sending out packets right away.  I’m not specifically familiar with the Xilinx board, but Ethernet MACs are all pretty similar in design.  I’d review the driver to ensure that the bit to mark the send buffer ready to send (Ready, Own, etc.) is set correctly when a packet is enqueued in the send queue, and that the hardware isn’t pending on something to be done.

 

---

From: address@hidden [mailto:address@hidden On Behalf Of Gary Olson
Sent: Friday, October 03, 2008 7:59 AM
To: address@hidden
Subject: [lwip-users] LWIP Delay between Consecutive Packets

 

Hi Kieran and Bill,

  I thought I had setup checksum offloading but I found that the software settings had to be set up as well as the hardware.  This did improve my bandwidth about 20 percent but still

To slow for my needs.  Still have delays between my UDP packets around 400 us.  My bandwidth for TCP transfers are around 70 – 80 Megabits per second.

 

I am using RAW API and a ML505 Xilinx Board with a Microblaze processor running at 100 Mhz and PLB 125 Mhz  I can send 3 Picture Frames in .8 seconds  one frame being 2 Million Bytes or 16 million bits.

 

3 * 16 Million bits = 54 million bits in .8 seconds  = 67.5 million bits per second using TCP.

 

Here is my settings:

 

BEGIN LIBRARY

 PARAMETER LIBRARY_NAME = lwip

 PARAMETER LIBRARY_VER = 3.00.a

 PARAMETER PROC_INSTANCE = microblaze_0

 PARAMETER tcp_mss = 4560

 PARAMETER ip_frag_max_mtu = 5500

 PARAMETER ip_reass_bufsize = 9000

 PARAMETER mem_size = 1000000

 PARAMETER pbuf_pool_bufsize = 4600

 PARAMETER tcp_wnd = 5560

 PARAMETER tcp_snd_buf = 25000

 PARAMETER tcp_tx_checksum_offload = true

 PARAMETER tcp_rx_checksum_offload = true

 PARAMETER pbuf_pool_size = 550

 PARAMETER memp_n_pbuf = 32

END

 

I have been using LWIP I think Version 3,  -llwip4 is my link against library command.

 

 

Additionally I played with tcp_wnd and tried raising it all the way to 60000 it might have helped some but brought it back down because

Some one said I should have the FIFO size for it in my hardware MAC.  I did notice that when I brought the tcp_wnd that the GAP started to

Vary from shorter to longer in a serial type of sequence.

 

 

Thank You,

Gary Olson