[HPADM] SUMMARY: Nkthreads questions

From: Jeff Cleverley (jeff.cleverley@avagotech.com)
Date: Wed Jan 09 2008 - 18:03:22 EST


Greetings,

I received replies from Denny Allain and Pal Baranyai. Some new spam
filters at work may have caused me to miss some other replies. Please
excuse me if I did not mention any others. The change of maxusers
mentioned by Denny was used and it bumped up the nkthread value almost
4x. We will put the new kernel into place at our next downtime window.
We are also putting in some other changes with recommendations found in
a Dave Olker book. Hopefully they will all work out.

On a side note, I had set up a script to put the new system file into
place and then build the new kernel. I received the following error:

kmupdate: Specified file /stand/build/vmunix_test is not kernel file.

After doing a google search, I got 1 hit listed below that fixed the
problem. There was another version of "nm" found in the path prior to
/usr/ccs/bin. Once I fixed that, the build script works fine.

http://forums12.itrc.hp.com/service/forums/questionanswer.do?admit=109447627+1199911155145+28353475&threadId=942498

I thought I'd mention this in the email and maybe I'll get 2 hits the
next time :-)

Thanks again,

Jeff
> Hi Jeff,
>
> make maxusers = 512 and max_thread_proc = 15 * num_clients
>
> there are other formulas that will change accordingly with those two
> changes and should get you in the ballpark with your server.
> In case you don't know it, maxusers has nothing to do with the number
> of users on the system, it's a seed value for the formulas that use
> it, it helps keep the other formulas in step with each other.
>
> Jeff Cleverley wrote:
>> Greetings,
>>
>> We did some changes to our physical network over the holidays, and
>> also corrected some nfs mount options for our linux workstations and
>> compute boxes. These changes have them talking to our hpux servers
>> (11.11, rp5470) via tcp instead of udp like they used to. Since we
>> made that change, I'm seeing the following messages in syslog on one
>> server:
>>
>> WARNING: tcpd_thread_create: thread_create failed: 12
>> kthread: table is full
>> WARNING: tcpd_thread_create: thread_create failed: 12
>>
>> Here are some kernel parameters (from kmtune) on that given server
>> that will probably be of interest:
>>
>> max_thread_proc 256 - 256
>> maxusers 128 - 128
>> nproc 1044 - (20+8*MAXUSERS)
>> max_thread_proc 256 - 256
>> nkthread 1843 - (((NPROC*7)/4)+16)
>>
>> I'm expecting that I'm going to have to grow some of these numbers to
>> get rid of the messages. From what I can tell, I have an average of
>> 1500 nfsktcpd threads running from the 1 process. This is from gpm
>> and a little script I found on the web. We have several hundred
>> linux compute boxes along with ~100 linux desktop boxes. According
>> to a Dave Olker book, the system allocates 10 threads per client
>> connection. I don't see how to break the threads out by what they
>> are actually doing.
>>
>> The 2 questions are:
>>
>> 1. Shouldn't the max_thread_proc limit nfsktcpd to 256 threads?
>>
>> 2. Is there a good way to tell how large I should make these
>> parameters? Should I just figure the 10 threads per client * number
>> of clients and make nkthreads bigger than that?
>>
>> Thanks,
>>
>> Jeff
>>

-- 
Jeff Cleverley
Unix Systems Administrator
4380 Ziegler Road
Fort Collins, Colorado 80525
970-288-4611
jeff.cleverley@avagotech.com
--
             ---> Please post QUESTIONS and SUMMARIES only!! <---
        To subscribe/unsubscribe to this list, contact majordomo@dutchworks.nl
       Name: hpux-admin@dutchworks.nl     Owner: owner-hpux-admin@dutchworks.nl
 
 Archives:  ftp.dutchworks.nl:/pub/digests/hpux-admin       (FTP, browse only)
            http://www.dutchworks.nl/htbin/hpsysadmin   (Web, browse & search)


This archive was generated by hypermail 2.1.7 : Sat Apr 12 2008 - 11:02:59 EDT