Select Loop and comm_remove_close_handler

Hi there,

I am experiencing a few strange (at least for me) effects:

First of all the setup:
 Sun Ultra 1 / 416 MB RAM
 16 GB cache_swap
 60 MB cache_mem
 1.NOVM.18 -DPOLL FD_SET to 4096
 Solaris 2.5.1 with rlim_fd_max=4096 and rlim_fd_cur=1024

1) We were running the cache for 3 days when squid died with the following
error message:

FATAL: comm_remove_close_handler: Handler not found!
 
 Squid Cache (Version 1.NOVM.18): Terminated abnormally.
 CPU Usage: user 20156 sys 28661
 Maximum Resident Size: 0 KB
 Page faults with physical i/o: 1095127
 97/12/04 15:03:19| storeWriteCleanLog: Starting...

Did anybody experience anything like that?

2) Whenever we restart/start up squid the cache manager reports a time for
the select loop of up to 300 ms and as the load of the cache slows down in
the evening, this time drops to about 12ms. But the strange thing is: when
the load picks up the following morning, the time STAYS at 12 to 17 ms
until the next restart.

3) Every time the number of filedescriptors reaches about 1024 the response
time (echoping) goes way up. Is there a certain parameter that has to be
tuned in Solaris or Squid to avoid this delay? (squid is compiled with
'poll', so more than 1024 descriptors are possible, but it's so slow :-(

Thanks for any comments,

  Michael.

-- 
Michael Rommel, ATD TD5 NT, Siemens AG, Erlangen, Germany