[Click] About memory allocation

[Guido Alejandro Gavilanes]

Hello Eddie and all developers which make this project a very useful
tool for research!!

I have been working on a switch prototype using click (1.5.0) using
ethernet encapsulation and forwarding between 2 interfaces on a host.
Trying with some ellaborated configurations (after having tried a bit
with click elements functionality), I began to perform some gross
testing with certain amount of packets per second, and then some
problem started to appear...

- I am using 6000 packets/sec on one interface and going out from the
second interface. Packets are pings and are generated by a tester
machine with two NICs. Those pings generate a throughput of 18Mbps
according to my tester. This traffic is unidirectional and just
forwarded by my click device.

- internally I use an Etherswitch to enable the host to perform
learning and use some other interfaces later. It is running on Kernel
mode and not running anything else or elements different from the
distribution.

- in this test, The host running the click configuration runs out of
memory in a matter of few seconds, more or less it looses 3Mbytes of
memory per second (I tested this using command cat /proc/meminfo  ->
MemFree  information). In the console, it appears this message

" __alloc_pages: 0-order allocation failed (gfp=0x...)",

then it starts killing processes and finally crashes, having to
reboot, or putting a script to monitor memory and shutting down click
when a limit is reached. I have seen on linux pages this is certainly
a memory problem, but anyway after unloading click module, memory is
not freed.

- When I replaced the etherswitch element with another void internally
hard-connected element (due to the fact etherswitch is experimental),
it also looses memory at a lower rate, (more or less 3KBytes/sec) with
the same packet rate described above. (I tested it for 1000 seconds
and noted the continuously decreasing memory)

- The same occurs by just replacing this configuration by a simpler
one, just receiving from an eth interface and putting through an
output queue to the other interface.

- I searched in the mailing lists, but I dont see "memory leaks"
problems (in fact just one
https://amsterdam.lcs.mit.edu/pipermail/click/2006-June/005022.html),
and that's why I guess my problem may require your attention.

-> The host machine I am using is a pentium 3, @800 MHz, 128MBytes in
RAM, Swap  256.
-> Operating System: Red Hat 9.0, running a kernel downloaded from
www.kernel.org ( version 2.4.32) and patched accordingly to click
instructions.

Ok, this is the problem, I would like to read some ideas around it,
and Thank you in advance for any additional information!

Guido