[Discussion] Unable to set X_{R|S}BufMode

[Richard Carlson <rcarlson@xxxxxxxxxxxxx>]

Gang;

At last weeks SC conference I found that one of the NDT sever at the 
Abilene New York PoP had very asymmetric throughput.  You could get 100 
Mbps from a server to the client but only 7 Mbps in the opposite 
direction.  I finally traced it down to a configuration issue in the NDT 
server.  The server has the Web100_autotuning turned off 
(/proc/sys/net/ipv4/web100_{r|s}bufmode = 0).

I manually turned them on and achieved 400 Mbps with a client in the 
Seattle PoP.  Instead of leaving them on, I decided that my code should 
check and enable the auto-tuning code on a per-connection basis.  This 
would ensure that the NDT server is going to operate properly, and I can 
generate a log entry if the server changes it so the admin can fix it if 
desired.

So I looked at the tcp_KIS.txt file and found that the variables are 
called  X_SBufMode and X_RBufMode in the 1.4 user lib.  I also verified 
that they were listed as read-write vars, so I could make the change on a 
per-connection basis.

I added the code to the NDT server and started testing.  I kept getting 
"web100_raw_read: fclose: permission denied" messages every time I tried to 
write a new value with the web100_raw_write() function.  I finally started 
using the readvar/writervar routines and found the same problem.

This lead me back to the userland lib code and I found a typo on line 1028, 
the perror message says web100_raw_read instead of web100_raw_write.

This lead me to look at the kernel patch code and I found that the 
X_{R|S}BufMode varialbes are listed as RO using the ADD_RO_STATSVAR() 
routine.  I also saw that the sndbuf and rcvbuf variables are listed as R/W 
using the ADD_RW_SKVAR() routine.

Am I correct in assuming these are bugs in the Web100 code?  Should I 
simply upgrade to a later patch/library or can I fix these myself?

Regards;
Rich



_______________________________________________
Discussion mailing list
Discussion@xxxxxxxxxx
http://internal.web100.org/mailman/listinfo/discussion