On Wed, Aug 26, 2009 at 08:38:42AM +0200, Simon Kallweit wrote:
> Sergei Gavrikov wrote:
>> I retried nc master/slave test with your template and suggested pbuf
>> values. For synthetic (I used tap interface) I got good results with
>> lwIP, but, for real target, nc test passed for 100% master load and no
>> load of slave side only. I got about 2Mbits per second with the DM900
>> Ethernet driver (that driver is too slow, it uses memcpy() on every 4
>> bytes arrived or sent), and my board gives only about 15 VAX Mips for
^^^^^^^^^^^
I checked, I got 11 Mips on the board.
>> RAM startup, and it seemed for me that was normal result.
>
> Did you use the old lwip port on that board too? Did you get similar
> results in performance?
Unfortunately, it is not possible run old nc_test_slave from the box,
because it depends on SO_REUSE, and we cannot manage SO_REUSE from CDL,
more that, if we force it, we'll get error
include/lwip/opt.h:353
Well, I forced SO_REUSE=1 and built the test. I did not noticed that I
got more Mbits with old lwip. The packets out of sequence often, but, at
the least the slave side was responsible with a loading.
FYI
old "udpecho" is responsible on
(while : ; do echo `date`; sleep 1; done)|nc -u -4 <board_ip> 7
and
(while : ; do echo `date`; sleep 1; done)|socat - udp4:<board_ip>:7
but could not manage stress "yes|".
old "tcpecho" could answer on stress "yes".
Simon, thanks for the feedback and the port. It will be interesting to
know other results from other sources (hardware).
Regards,
Sergei
