Re: [squid-users] RE: SUN E4000 - comm_accept and http_accept: accept_failure - Answers and Questions

It looks like this tuning has been done for an
earlier (most likely solaris 2.6) kernel.

If you _are_ using solaris 8 as mentioned below
(i missed the earlier part of the thread):

- *remove* the set priority_paging=1, and initially
  at least the other vm tuning (lotsfree etc). Solaris
  8 has a new paging algorithm (cyclic paging) by
  default which supercedes priority paging. Turning
  on priority paging disables the more advanced cyclic
  paging methodology.

- run 'kstat hme' or qfe, or whatever interfaces you
  are using. Alternatively run 'netstat -k' and look
  for the hme/qfe section. Check for any errors such
  as under the 'nocanput' field.

- run netstat -sP tcp; probably the most relevant field
  here is tcpListenDrop, but look for anything else that
  may look strange.

There's some good basic info about this kind of thing at:
http://www.sean.de/Solaris/tune.html

If your tcp stack is stuffed, you can tune it to cope,
or almost cope with your high watermarks, but in the
longterm you may want to look at spreading the load
over more network interfaces [maybe have a look at
Solaris 8 IP multipathing]

- Mike

> Hi All,
>
> Thanks for all...,
>
> Mail from: Adam Carter
>
> > Have you tweaked solaris's kernel in the same way on both boxes?
> > Solaris 8
> > has tons of kernel parameters you can modify, perhaps you are
> > running out of
> > some resource (check the logs), or just not optimally tuned. I know
> > the
> > default config left many of our busier boxes starved. None of these
> > were
> > dedicated squid boxes, but its still worth investigating. Let me
> > know if you
> > want more info on this and i'll dig up the configs.
> > Adam
> >
> This is the tune that has in the e4000 box...
>
> set rlim_fd_cur=65535
> set rlim_fd_max=65535
> *
> *
> * Tunning for squid diskd
> set msgsys:msginfo_msgmax=2049
> set msgsys:msginfo_msgmnb=8192
> set msgsys:msginfo_msgssz=64
> set msgsys:msginfo_msgtql=1024
> *
> *
> * Tunning for Solaris 8 - From Sun Microsystems
> set shmsys:shminfo_shmmax=524288000
> set shmsys:shminfo_shmmin=1
> set shmsys:shminfo_shmmni=100
> set shmsys:shminfo_shmseg=10
> set semsys:seminfo_semmnu=50
> set semsys:seminfo_semmap=100
> *
> set semsys:seminfo_semmni=200
> set semsys:seminfo_semmns=1000
> set semsys:seminfo_semmsl=100
> set msgsys:msginfo_msgmni=50
> set lwp_default_stksize=0x4000
> set rpcmod:svc_run_stksize=0x4000
> *
> set priority_paging=1
> set cachefree=191514
> set lotsfree=95757
> set desfree=47878
> set minfree=23939
> set fastscan=95757
> set slowscan=9575
> set handspreadpages=95757
> set maxpgio=256
>
> TCP TUNE...
>
> /usr/sbin/ndd -set /dev/tcp tcp_conn_hash 262144
> /usr/sbin/ndd -set /dev/ge adv_1000autoneg_cap 0
> /usr/sbin/ndd -set /dev/ge adv_1000fdx_cap 1
> /usr/sbin/ndd -set /dev/ge adv_1000hdx_cap 0
> /usr/sbin/ndd -set /dev/tcp tcp_xmit_hiwat 65535
> /usr/sbin/ndd -set /dev/tcp tcp_recv_hiwat 65535
> /usr/sbin/ndd -set /dev/tcp tcp_cwnd_max 65535
> #
> /usr/sbin/ndd -set /dev/tcp tcp_slow_start_initial 2
> /usr/sbin/ndd -set /dev/tcp tcp_time_wait_interval 60000
> /usr/sbin/ndd -set /dev/tcp tcp_conn_req_max_q 1024
> /usr/sbin/ndd -set /dev/tcp tcp_conn_req_max_q0 131072
> /usr/sbin/ndd -set /dev/tcp tcp_wscale_always 1
> /usr/sbin/ndd -set /dev/tcp tcp_max_buf 504857600/usr/sbin/ndd -set
> /dev/tcp tcp_smallest_anon_port 8192
> /usr/sbin/ndd -set /dev/udp udp_smallest_anon_port 8192
> /usr/sbin/ndd -set /dev/tcp tcp_rexmit_interval_min 2000 # 200 for
> laboratories
> /usr/sbin/ndd -set /dev/tcp tcp_ip_abort_interval 600000 # 10 min
> before drop
> /usr/sbin/ndd -set /dev/tcp tcp_ip_abort_cinterval 60000 # 60 sec
> to estab. conn.
> /usr/sbin/ndd -set /dev/tcp tcp_fin_wait_2_flush_interval 67500
>
> Mail from: joelja@darkwing.uoregon.edu
>
> Given the fact that the main squid process is only gonna sit on a
> single
> cpu I'd say an e4000 might be a little oddly sized as squid box... if
> the
> e250 had the 300 or 400mhz ultra IIi cpu's I could see it
> outperforming an
> e4000 even though the larger machine has a lot more cpu cache, given
> the
> amount of ram you have and the fact that e4000 have fibre channel
> controllers built-in I'd be inclinded to aim for a 100GB or so of
> cache
> dir. squeezing more performance out of the box might also involve
> running
> more than one instance of squid, and point some of the clients at one
> and
> some at the other...
> I understand..., but I have some questions:
>
> Do you mean: Running one instance of Squid per CPU?
> Do you mean: An e250 is better than a e4000 to run Squid?
>
> Mail from: vanzyla@cedara.kzntl.gov.za
>
> What kind of load does the server carry?
>
> Only run squid with user authentication...
>
> How many file descriptors are configured on your box?
>
> In Solaris Operation System are running with 65535 and in the squid
> are running with 32768 (I think that are the max value for squid...).
>
> AD
>
> __________________________________________________________________________________________________________________
>
> Do you have any idea to solve this problem ?
>
> Thanks in advance
>
> Best Regards
>
> --
> _________________________________________
> Hederson Velasco Ramos
> Technology Consultant
> Email: hederson.ramos@edb.ericsson.se
> EIN - Network Services - Ericsson - EDB

-- 
Michael Kiernan
Onet.pl S.A.
Krakow, Poland
http://www.onet.pl/