Bind 8.1.2 in.named memory leak in Solaris 7

[John A. Stewart]

In <14483.939389314 at kludge.mpn.cp.philips.com> Jim Reid <jim at mpn.cp.philips.com> writes:

>>>>>> "Joaquim" == Joaquim Eudes Mendes Gomide <jgomide at bancobrasil.com.br> writes:

>    Joaquim> I got an article on Sunsolve that says: "The DNS process
>    Joaquim> 4in.named4 continually consumes memory until no
>    Joaquim> memory is available. At this point in.named terminates
>    Joaquim> leaving a possibly large core file."  Is this problem a
>    Joaquim> Bind 8.1.2 or a Solaris version of Bind bug? The same
>    Joaquim> article says that there is no work around, too.  Does the
>    Joaquim> Bind 8.2 solves this problem?

>The article is a little bit misleading. It gives the impression that
>name servers guzzle memory until the system runs out and then named
>crashes and dumps core. This isn't quite true. It is true that the
>name server will die if it cannot get more memory from the OS. However
>that should be a rare event. And if your system can't let the name
>server process have more memory, then you have to contend with more
>serious underlying problems. [Like a grossly overloaded computer or a
>woefully inadequate local DNS infrastructure.]

>The size of a name server process grows as it receives queries. These
>make named cache more and more data as it looks up names and sends
>queries to other name servers. However the size of the process tends
>to stabilise after the name server has been running for a few
>days. Old entries get expired from the cache and the lookup patterns
>from local resolvers tend to request the same names.

>This "problem" is the way name servers work. They take more memory
>from the OS as they need to cache resource records. In general this
>doesn't grow without limit. I believe that BIND9 will behave more
>gracefully when named's memory allocation requests fail. However this
>isn't due to be released until next year and it may be a year after
>that before vendors ship it with their OS releases.

>FWIW the name server on a Solaris box here has been running for a few
>weeks now and it only uses 1.5M of VM. This isn't a noticeable load on
>the system and is nowhere near the resource limits of the OS.

   That's a good overview of how a dns name server should be working, however
there really is a big memory leak in the 8.1.2 version of bind that ships
with Solaris 7.  After migrating our primary dns server to Solaris 7 we 
were finding that it took less than a day to grow to more than 50MB in size!
A cron job which checks the size of the in.named process was restarting the
daemon at least once a day.  I worked around that problem by installing the 
8.2.1 version of bind.  It now takes several weeks before it hits that 
level of memory consumption.
-- 
John Stewart -- Computing and Communications Services, Carleton University
Internet: jstewart at ccs.carleton.ca                       613-520-2600x3707
"Why are so many people trying to get ahead when they already have one?"