Processors

Processors forum

State Resolved
Locked Locked
Replies 3 replies
Subscribers 99 subscribers
Views 642 views
Users 0 members are here

Support feedback

Options

Options

Related

Ethernet performance SysBios+NDK

dohan tang

Prodigy 60 points

Other Parts Discussed in Thread: SYSBIOS

Hi,

My platform is DM8148+SysBios+NDK, I created a UDP application to send out data(2M Bytes) to PC, each UDP packet with 1K bytes data block to avoid IP fragmentation.

The total packets sent will be 2000. I use wireshark to monitor the Ethernet raw frames, 2000 frames spends about 1.5 seconds. The interval between packets is about 0.8ms. How the improve the Ethernet send performance?

Thank you.

over 12 years ago

0 Alan DeMars over 12 years ago

TI__Mastermind 30830 points

If I calculate correctly, your throughput is around 11Mbps.

Thats pretty low compared to the benchmarks shown here:

http://www.ti.com/lit/an/spraaq5a/spraaq5a.pdf

The graphs show that performance decreases depending on CPU load. Do you have other threads that are consuming a lot of bandwidth?

Have you tried the recommendation mentioned here:

http://processors.wiki.ti.com/index.php?title=Network_Developers_Kit_FAQ#Q:_I_wonder_if_I_could_extract_some_additional_performance_from_NDK._Are_there_any_tips_to_make_sure_my_system_is_performing_at_top_notch.3F

Alan

0 dohan tang over 12 years ago in reply to Alan DeMars

Prodigy 60 points

Dear Alan,

First, You are right, the performance is around 11Mbps.

You provide the link about EMAC performance which controlled by DSP C67xx.
But, in my platform, EMAC is controlled by Cortex-A8 on DM8148, not C674x.

So how to improve the performance?

Second, in spraaq5a.pdf, the performance is measured by "non-copy" method, how about "copy" method?

What difference between these two method?

Thaank you.

0 Prasad Joshi over 12 years ago in reply to dohan tang

TI__Intellectual 1010 points

Hello Dohan,

non-copy method uses pointer passing method for data buffer.

copy method uses memcopy for data buffer copying inside the NDK stack.

Therefore it is always advisable to use non-copy APIs for better performance.

The other recommendation for better throughput is to use internal data memory in place of external data memory.

This will save CPU cycles for Cache flushing too.

Please have a look on NDK user guide (spru523_ug.pdf) Section 3.1.3 (NDK Memory Sections) for further details.

This document comes with NDK package.

Regards

Prasad V. Joshi