Re: fix(?) for Solaris 2.5, running out of filedescriptors.

>Date: Thu, 28 Nov 1996 16:40:18 +1100
>From: Anthony Baxter <arb@connect.com.au>
>Subject: fix(?) for Solaris 2.5, running out of filedescriptors.
>To: squid-users@nlanr.net
>
>Hi folks -
>
>I've got a fix/workaround for getting squid to use more than 1024 fds
>on Solaris 2.5. The problem is with the select() call, which, in Solaris,
>is just a wrapper around poll(). I wrote my own select() call (which calls
>poll()) and set it to use up to 4096 filedescriptors. According to Sun,
>various kernel thingys go byebyes when you push it any higher than that.

Hmm. When we found we needed more than 1024 descriptors on a couple of Solaris
2.4/2.5 news feeder boxes, we just added the following lines to /etc/system
and rebooted...

* set hard limit on filedescriptors
set rlim_fd_max = 2048
*
* set soft limit on filedescriptors
set rlim_fd_cur = 4096

Related entries that may also help performance for some applications (like
INN) are

* increase directory name lookup cache (per Sun Performance and Tuning,
* aka the "Porche Book", p. 190)
* ufs_ninode=ncsize=maxusers*17+90 = 10970
* we've been running with 20,000 for a while
set ncsize = 20000
*
* increase inode cache (per Sun Performance and Tuning, p. 192)
set ufs_ninode = 40000

Note: closely investigate the suitability of these changes for your own
environment; these are examples of what work for us, not a recommendation
or recipe to be blindly applied. Your mileage will vary, caveat emptor, etc.

Regards,

Joe St Sauver (joe@oregon.uoregon.edu)
University of Oregon Computing Center
Received on Thu Nov 28 1996 - 10:35:36 MST