Hi Karl, all

On 10/15/07, Karl Czajkowski <[EMAIL PROTECTED]> wrote:
> When this is going on, how many connections are showing up under
> netstat, i.e. in various TIMED_WAIT states?  Is it possible that your
> pool of 1000 sockets is being exhausted and leading to more rapid
> re-use of existing sockets?

No, during these tests I have been the soul user of the machine. It is
a vmware 64bit CentOS 5 guest running in a Windows Server 2003 host
that I have been configuring as the grid front end for a recently
acquired cluster.

> A little googling on socket reuse revealed some problems discussed
> with web and POP mail servers where source socket addresses were being
> reused while a server still had a TIME_WAIT entry for the same address
> from an older connection (on the same host pair)...

I have previously looked at what was going on using netstat and only
saw globus connections in the 43000:44000 range.

> It might help for others if you mentioned the operating system(s) and
> version(s) of the endpoints in question (in order to understand if there
> are any known TCP/IP quirks there).

Yes, sorry I meant to add that but hit send in a hurry. I mentioned
the server OS above, the server is running GT4.0.5 (I followed the
quickstart). The client machine is a Fedora Core 5 box with GT4.0.2 WS
clients. The iptables on the server is up to date with the current
repos.

> As I think you are looking at packets created by the TCP stack on the
> OS and not packets crafted by Globus Toolkit software, I wonder if you
> are really just seeing an underlying OS bug or misconfiguration that
> is being tickled due to, e.g., fast connection rates with a certain
> file staging workload.

It is certainly a possibility that this is an artifact being
introduced elsewhere in the communications, it could be a bug in the
netfilter state tracking code. However, as I mentioned I have never
had any other trouble with the state module and I figure it would be a
more thoroughly exercised bit of code than globus. I guess I could
test that reasonably easily by tcpdump-ing the client activity too and
matching the outgoing packets there, however I wouldn't expect to find
any differences as this is all on the input chain and nothing to do
with any nat transformations.

I'm happy to provide for logs/data if you can think of anything that
would help shed more light.

Cheers,
-Blair

-- 
In science one tries to tell people, in such a way
as to be understood by everyone, something that
no one ever knew before. But in poetry, it's the
exact opposite.
 - Paul Dirac

Reply via email to