Handling 100.000 packets/sec or more

Wed Jan 14 12:15:51 PST 2004

Adrian Penisoara <ady at freebsd.ady.ro> writes:

|Hi,
|
|  At one site that I administer we have a gateway server which services
|a large SOHO LAN (more than 300 stations) and I'm facing a serious
|issue: very often we see strong spoofed floods (variable source IP and
|port, variable destination IP, destination port 80) which can go as far
|as 100 000 packets/sec!
|
|  Of course, the server (FreeBSD 5.2-REL, PIII 733Mhz, 256Mb RAM, 3COM
|3C905B-TX aka xl0 with checksum offloading support) has a hard time
|swallowing this kind of traffic. The main issue are the IRQ interrupts:
|over 15000 interrupts/sec which consume more than 90% of the CPU time.
|We got ingress filtering so the packets go no further than the firewall
|(which, BTW, is not the issue, even disabling it it's the same
|problem). The system is still responsive but the load average goes as
|high as 10 and the interface is losing packets (input errors) which
|dramatically affects legitimate traffic, besides mbuf(9) starvation. We
|are taking down the culprit clients, but this takes time and we need
|the other clients not to be affected by it.
|
|  What can I do to make the system better handle this kind of traffic ?
|Could device polling(8) or just increasing the kernel frequency clock
|to 1000Hz or more improve the situation ?
|  What kind of network cards could face a lot better this burden ? Are
|there any other solutions ?
|
	Try fxp. It has better polling support, and there's the advantage of
the link0 flag. When it's set, the interface won't send interrupts to
the kernel for each packet it catches from the wire, but instead will
wait until its own buffer is full, and generate an interrupt afterwards.
It should be a great deal of improvement when asociated with device
polling. As you surely know, when the kernel receives an interrupt from
an interface, it masks all further interrupts and schedules a polling
task instead.
	In matters of hardware improvements, I can only think of upgrading to a
system board with a higher bandwidth between the south and the north
bridges, and of course, using a CPU that has a higher clock rate.

|  On a side note: what would be a adequate formula to calculate the
|NMBCLUSTERS and MBUFS we should set on this server (via boot-time
|kern.ipc.nmbclusters and kern.ipc.nmbufs) ?
|

	I'm still thinking about that ...

----
If it's there, and you can see it, it's real.
If it's not there, and you can see it, it's virtual.
If it's there, and you can't see it, it's transparent.
If it's not there, and you can't see it, you erased it.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 187 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-net/attachments/20040114/50127499/attachment.bin