HP-UX TCP Stack Limits to Obey?

[Ralph Grothe]

Hello,

I need to find out about any possible "limitations" or performance degradataions of the hpux tcp stack, and if they exist how to overcome them maybe by kernel or ndd tuning, or patching.

I was told that the application developers of one of our customers threat to release a new version of their application software which as rumour has it seems to be heavily tomcat/mod_jk based.
This shift to Java lets me fear the worst for the performance.
I was also told that it is roughly estimated that the peaks of concurrently established sockets of client connections may raise to well beyond 2000.
(I wonder how, if their client base hasn't risen at the same time?)
Currently on an average working day we are below 2000 sockets.
(if you are interested, please have a deko at the rrdgraphs I attached to this posting)
As far as I can remember from former threads of this kind I dared to plague you with,
some network gurus like Rick Jones et al. mentioned that there were no such inherent restrictions in the network stack, but only the system's resources such as cpu and memory.
But anyway, I wanted to make sure if there couldn't be done anything from my sysadmin part to make it a smooth transit.
Maybe someone of you has also experience with resource hunger and stability of tomcat applications on hpux?

Regards
Ralph

[IT_2007]

Ralph,

You can put limit of tcp socket connections adding entry in nddconf file.

transport_name[0]=tcp
ndd_name[0]=tcp_time_wait_interval
ndd_value[0] =3000

[Chan 007]

Ralph,

PLease check the link below for NDD info,

ftp://ftp.cup.hp.com/dist/networking/briefs/annotated_ndd.txt

But I recon, that you should look this ptoblem from application side - TOMCAT or your coding also. Check for the parameters that are recomended for tomcat.

Chan

[Ralph Grothe]

Thank you both,

actually, I think this is the link to Rick Jones'es annotations which I have bookmarked some time ago.
I wonder I he has written an update by now?

[rick jones · Accepted solution]

I would strongly suggests NOT to set the tcp_time_wait_interval to 3 seconds (3000 milliseconds). The HP-UX TCP stack can more than easily deal with thousands and thousands of TIME_WAIT connections. The default at 60000 (one minute) should be more than sufficient.

I've not updated my ndd writeups. Officially I'm not in the UX organization any longer, so my "day job" incentive to update that has waned a bit :) Still, I do play with UX boxen from time to time, and I like telling other people what to do :) so I may go ahead and try for an update.

[Ralph Grothe]

Hello Rick,

I'd like to tell you how much I appreciate your nnd parms annotations.
They always are a valuable source of explanation of the obscure hpux network voodoo.
Therefore, I'd like to encourage you to keep on with occassional revisions should changes in the stack's ndd interface necessitate and justify this effort.

[rick jones]

I'll see what I can do. FWIW I thought that TPTB had already taken what I'd written and massaged it into a more official form - probably in the knowledge base somewhere?