What settings for a large load?

[Andrey Savochkin]

On Fri, May 05, 2000 at 08:12:53AM -0700, Kevin Mills wrote:
> > I doubt that it's a driver problem.
> > Check more careful what happens on other layers of the system.
> > Doesn't the kernel consider your traffic as a SYN-flood attack?
> > Check by tcpdump why connections are reset.  If really the client
> > sends RST
> > packet?
> 
> What would tell me if Linux considers itself under attack?  Are there

You would see corresponding messages in your log.

> settings in the kernel I should tweak?  I don't see any kernel messages in
> my logs indicating anything.  And why would SunOS be happy with the same
> simulation?  I can certainly turn on my sniffer, but it is hard to read it

If SunOS had the same Linux kernel, the difference between Linux and SunOS
would be a surprise :-)
In reality, there is an enormous amount of differences in the kernels which
can cause the performance difference.

> with so much traffic :)

Read only interesting places.
Like packets with RST bits and nearby.

> 
> > Check you client, too.  Doesn't it do some insane things?
> > E.g. does it really calls connect(), sends/receives the data and
> > closes the
> > connection?  Or works just on packet level, thus, likely, disabling flow
> > control etc?
> 
> The clients are doing normal socket operations.  Each client calls
> connect(), send(), recv() and then pauses 250 milliseconds.  It then does 20
> more send()/recv() calls and then disconnects, pauses 250 milliseconds and
> starts over at the top with a connect().  Not too insane, I don't think.

The socket is TCP, isn't it?

> 
> > Your situation isn't a matter of settings.  It a problem somewhere.
> 
> Again, my Sun box has no troubles with the simulation so it must be a matter
> of settings *somewhere*.  For instance, I was receiving 'eth0 reports no
> more resources' from the driver until I increased the FIFO settings.

OK, if you're insisting on setting issue, 'card reports no resources' means
that your system temporary run short of free memory.  Increase
/proc/sys/vm/freepages numbers.

> Perhaps the issue isn't all at the driver level, but I would expect Linux to
> be able to handle this load; wouldn't you?

I expect, too.
Dig out further.  Make your dumps and netstat output publicly available, may
be somebody gets the idea.

Best regards
					Andrey V.
					Savochkin
-------------------------------------------------------------------
To unsubscribe send a message body containing "unsubscribe"
to linux-eepro100-request@beowulf.org