[Beowulf] always raw

050675 at student.unife.it 050675 at student.unife.it
Wed Sep 1 03:59:44 PDT 2004


Hi all,
I'm still doing my experiments with raw ethernet, but something strange
happens. I've attached to this mail my two test programs. They are very
simple but useful for me to test the performance of a tcp connection. In
particular, as you can see from the source code, I send 1 million tcp
packets composed by an unsigned char buffer of TCP_MSS (512) + 40 bytes.
The 40 bytes represent the ip and tcp headers. I send data with write()
and I receive them with read(). Running my two programs produces the
following results to the receiver node (where tcpserver runs)

Bytes received: 552000000
Time elapsed: 5.7 sec
Throughput: 91 Mbytes/sec

Now, the same experiment reproduced with raw ethernet illustrates a
throughput of 87 Mbytes/sec and the time elapsed is 6.3 sec. How is it
possible? Raw ethernet doesn't need retransmission, ack, etc. like tcp,
but here tcp seems to be the best. I don't understand.
Probably there's something wrotten in my programs (so excuse me if I post
them, but I don't know how to do at this point).
Finally, I'm studying the problem of packet loss with raw ethernet, as
I've told in my past mails and I'm having good results by chancing the
receive buffer size and some parameters of the nics driver. Stay tuned and
I'll post other info if interested.

Thanks for yours help

Simone Saravalli

-------------- next part --------------
A non-text attachment was scrubbed...
Name: tcpclient.c
Type: text/x-csrc
Size: 3506 bytes
Desc: not available
Url : http://www.scyld.com/pipermail/beowulf/attachments/20040901/f7b090ed/tcpclient.bin
-------------- next part --------------
A non-text attachment was scrubbed...
Name: tcpserver.c
Type: text/x-csrc
Size: 3368 bytes
Desc: not available
Url : http://www.scyld.com/pipermail/beowulf/attachments/20040901/f7b090ed/tcpserver.bin


More information about the Beowulf mailing list