Neeu
asked on
UDP Full Sockets
Hi,
We increased the network speed from 10Mbit to 100Mbit and we experiencing lost application packets at our communication sessions between client’s side and our communication application that run at the DS20 Tru64 4.0F UNIX server.
We check “netstat –I ee0” and we don’t have any errors at the server network interface, but when we checked the netstat option using “netstat -p udp” we got that “full sockets” is not zero.
The output is like the following:
udp:
12003 packets sent
48193 packets received
0 incomplete headers
0 bad data length fields
0 bad checksums
2170 full sockets
12943 for no port (12916 broadcasts, 0 multicasts)
At the documentation and we saw that:
“When to Tune
Use the netstat -p udp command to check for full sockets. If the output shows many full sockets, increase the value of the udp_recvspace attribute. “
Our questions are:
1. Can it be related to the communication problem I describe above?
2. How “full sockets” effects our system?
3. The udp_recvspace set to 41600 , if we should take the advise from the documentation what should be the proper setup ?
4. What is the effect of increasing the udp_recvspace parameter ?
Please Advise,
Ran.
We increased the network speed from 10Mbit to 100Mbit and we experiencing lost application packets at our communication sessions between client’s side and our communication application that run at the DS20 Tru64 4.0F UNIX server.
We check “netstat –I ee0” and we don’t have any errors at the server network interface, but when we checked the netstat option using “netstat -p udp” we got that “full sockets” is not zero.
The output is like the following:
udp:
12003 packets sent
48193 packets received
0 incomplete headers
0 bad data length fields
0 bad checksums
2170 full sockets
12943 for no port (12916 broadcasts, 0 multicasts)
At the documentation and we saw that:
“When to Tune
Use the netstat -p udp command to check for full sockets. If the output shows many full sockets, increase the value of the udp_recvspace attribute. “
Our questions are:
1. Can it be related to the communication problem I describe above?
2. How “full sockets” effects our system?
3. The udp_recvspace set to 41600 , if we should take the advise from the documentation what should be the proper setup ?
4. What is the effect of increasing the udp_recvspace parameter ?
Please Advise,
Ran.
ASKER
I used "netstat -i" and "netstat -s", there is no new information, there is no Input or Output errors at the network interfaces.
Then host is full duplex and switch is half duplex
ASKER
How can you say that ?
Both host and switch are Full Doplex - no I/O errors at the network recorded.
Both host and switch are Full Doplex - no I/O errors at the network recorded.
Did you look at switch LEDs ???
Host is full duplex, so it receives while sending
Switch is half duplex, so it drops while sending.
No counters on the host shows this, try pinging something stable from host, like your workstations, if I am right there should be huge losses
Host is full duplex, so it receives while sending
Switch is half duplex, so it drops while sending.
No counters on the host shows this, try pinging something stable from host, like your workstations, if I am right there should be huge losses
ASKER
At the UNIX we use at /etc/inet.local the follwoing command in order to fix the NIC configuration :
lan_config -i ee0 -a 0 -s 100 -x 1
As I said there is no errors between Unix to the Switch, I duble checked the configuration and setup both wotking 100Mbit FD.
lan_config -i ee0 -a 0 -s 100 -x 1
As I said there is no errors between Unix to the Switch, I duble checked the configuration and setup both wotking 100Mbit FD.
look at netstat -na
Proto RecvQ SendQ are first columns
If RecvQ is not 0 - server process lags
If SendQ - network or another communication end is slow
/
recvspace is of little help if receiving process is overloaded.
/
"full socket" is full socket buffer, for receiver program to process.
Proto RecvQ SendQ are first columns
If RecvQ is not 0 - server process lags
If SendQ - network or another communication end is slow
/
recvspace is of little help if receiving process is overloaded.
/
"full socket" is full socket buffer, for receiver program to process.
ASKER
Any ideas ?
do not rise socket buffer, it simply hides performance problems...
The idea is to adjust (lower) program priorities (aka renice) so they work friendlier, without overloading buffers, i.e lower priority for programm filling socket buffers.
ASKER
The problem is that this is a client server program, so I don't have real control on the client side that send requests.
Look at netstat - whether problem with full buffers is in application receiving requests or sending responses back so you will heave clear picture whether you need to tune server application or network connection. (if you have slow links and full send buffer - that is expected, only in this case larger buffers may help, and only if full socket send buffer blocks server application, otherwise it is just a small but waste of memory.
If input buffer is full, you need to tune your application for better performance/parallelism (not rise priority)
since I do not know which problem you encounter I cannot help more.
If input buffer is full, you need to tune your application for better performance/parallelism (not rise priority)
since I do not know which problem you encounter I cannot help more.
ASKER
Thanks for now for your help, we are now in contact with HP (Compaq) experts, the problems I mention above also happen with TCP sessions, and that's very odd seem like a problem with the OS.
The Unix OS not transferring packets trough or losing packets in the way, we saw that application response from the server although they are logged at the applications logs files not sent back to client side.
The Unix OS not transferring packets trough or losing packets in the way, we saw that application response from the server although they are logged at the applications logs files not sent back to client side.
ASKER CERTIFIED SOLUTION
membership
This solution is only available to members.
To access this solution, you must be a member of Experts Exchange.
ASKER
After installing the latest Patch kit we over came the problem and now the full sockets is 0 , still the application problem didn't resolve yet.
SOLUTION
membership
This solution is only available to members.
To access this solution, you must be a member of Experts Exchange.
PAQed with points refunded (300)
modulo
Community Support Moderator
modulo
Community Support Moderator
ASKER
I should add we found the source of problem, it was related to the DE600 Network interface that have a low signal and was replaced by DE500 that recover this problem.
We found that in DE600 the long cable about 80 Meters between ethernet switch and NIC couse the problem , altough it shoud be supported .
Thank you gheist for your support !
Ran.
We found that in DE600 the long cable about 80 Meters between ethernet switch and NIC couse the problem , altough it shoud be supported .
Thank you gheist for your support !
Ran.
kind of my first network gear suggestion ???
Neeu -
It might be fair if you asked to have the question re-opened so you can award points to gheist.
jmcg
EE Cleanup Volunteer
It might be fair if you asked to have the question re-opened so you can award points to gheist.
jmcg
EE Cleanup Volunteer
ASKER
I agree,
I tried to do that but I didn't find the way.
how I re-open the question ?
Ran.
I tried to do that but I didn't find the way.
how I re-open the question ?
Ran.
netstat -s
netstat -i
will show more significant statistics.