TMS320C6678: About TCP throughput Issue

Hi Yordan,

Yimin is my colleague.

For NDK TCP TX throughput testing, I try to do all suggestion as I can.

But the TCP TX throughput does not reach acceptable value when I send 200 bytes every time.

All that I have modified are:

1. Increase TCP buffer.
  Set CFGITEM_IP_SOCKTCPTXBUF to 64000
  Set CFGITEM_IP_SOCKTCPRXBUF to 64000
  Set CFGITEM_IP_SOCKTCPRXLIMIT to 64000

2. Adject NIMU TX/RX buffer ratio.
  Increase NIMU_NUM_TX_DESC up to 94u
  Decrease NIMU_NUM_RX_DESC down to 32u
  (in pdk\packages\ti\platform\resource_mgr.h)

3. Increase NDK frame buffer.
  Set PKT_NUM_FRAMEBUF to 960
  (in ndk\packages\ti\ndk\stack\pbm\pbm_data.c)

4. Locate NDK_PACKTMEM section on MSMCSRAM.
  Program.sectMap[".far:NDK_PACKETMEM"] = {loadSegment: "MSMCSRAM", loadAlign: 128};

5. Rebuild NDK as release mode and set size optimization from -ms2 to -ms1

6. Allow more TCP connection.
  Set RAW_PAGE_COUNT to 256
  (in ndk\packages\ti\ndk\os\mem_data.c)

7. Set NIMU RX accumulator.
int32_t Setup_Rx (NETIF_DEVICE* ptr_net_device)
{
    accCfg.timerLoadCount to 5;
    if ((result = Qmss_programAccumulator (Qmss_PdspId_PDSP1, &accCfg)) != QMSS_ACC_SOK)
}

According to the link (e2e.ti.com/.../283019)
TI Employee say : "From many customer experience with NDK in the e2e forum, the TCP throughput is about 600-700Mbps"

This is fine to us. we probably accept TX throughput on 600-700Mbps.

But the issue is that,
when I send a smaller packet size than 7*1460 every time, I got a smaller TX throughput too.
Our product care about TX latency therefore I have to send packet immediately.

When I just modify test packet size from (7 * 1460) to 200 bytes,
The TX throughput is down to 120Mbps.
It is almost 5 times lower than the reasonable throughput (600Mbps).

Furthermore, I observe every packet by Wireshark.
I find test peer's window size never been full.
Every outgoing packet does not very close.
I guess that I should modify somewhere NDK code to increase TX throughput.

Sending 200 bytes packet on general PC can reach 900Mbps easily.

Following are my test result:

a) C6678 send 10220 bytes every time.

TX throughput is 640 Mbps.

b) C6678 send 1024 bytes every time.

TX throughput is down to 360 Mbps.

c) C6678 send 200 bytes every time.

TX throughput is down to 120 Mbps.

d) General Linux PC send 200 bytes every time.

Linux TCP/IP stack do collecting outgoing packet job is well.

TX throughput is still 900 Mbps.

Any advice is fine.
Thank so much,

Sean Lin