[AusNOG] Tech Q: Looking for Cheapest 10G Switches for Hobby Lan.

[Mark Smith]

On Mon, 19 Jul 2010 11:30:54 +1000
"Sean K. Finn" <sean.finn at ozservers.com.au> wrote:

> The other thing to be aware of here:
> 
> Jumbo Frames increases latency, obviously.
> 
> Were you referring to L2 latency or L3 ?
> 

Layer 3 I suppose, but I'm not really sure it matters. Since it was a
measure of the time it took between when the packet entered the host,
was processed, and sent back out, and the processing was functionally
the same, I think the significance is that the 450Mhz 32 bit P3
with a PCI 100Mbps NIC was consistently faster at performing the same
processing as the 1.6Ghz (clocked down from 2.4Ghz) 64 bit Q6600 with a
PCIe 1Gbps NIC. I'll admit it wasn't very scientific, because my goal
wasn't specifically to measure the differences - it was to test the
code out. The significance of the difference was surprising though,
which is why it stood out. Up until that point I'd have assumed that
there was nothing a machine built in 1998, and a network card of a
similar vintage, would be able to do faster than a machine from 2008,
running the same code. Then again, even cheap (~$70) yet quality 1Gbps
cards are doing far more processing onboard, such as TCP Segmentation
Offload, 802.1Q VLAN tag stripping etc, and they all add processing
latency, even if the frames don't carry those protocols.



> S
> 
> 
> 
> -----Original Message-----
> From: ausnog-bounces at lists.ausnog.net [mailto:ausnog-bounces at lists.ausnog.net] On Behalf Of Mark Smith
> Sent: Monday, 19 July 2010 8:16 AM
> To: Narelle
> Cc: ausnog at lists.ausnog.net
> Subject: Re: [AusNOG] Tech Q: Looking for Cheapest 10G Switches for Hobby Lan.
> 
> On Sun, 18 Jul 2010 19:50:22 +1000
> Narelle <narellec at gmail.com> wrote:
> 
> > On Sat, Jul 17, 2010 at 10:20 AM, Mark Smith
> > <nanog at 85d5b20a518b8f6864949bd940457dc124746ddc.nosense.org> wrote:
> > > Want to make their head explode? In some experimentation a while back, I
> > > measured latency of packet processing of an old Netgear FA312 100Mbps
> > > NIC in a P3 450Mhz, verses an Intel 1Gbps PCIe in a Q6600 quad core. The
> > > Netgear had lower latency, and IIRC, significantly lower latency. Of
> > > course, all those measurements were in microseconds, and therefore
> > > probably irrelevant to actual gaming, but I don't think they actually
> > > care. The length of the cable was measurable too at that scale. (If you
> > > get pictures of a fight breaking out at a LAN party over old 100Mbps
> > > NICs, and who's sitting closest to the switch, send them to me
> > > please :-) )
> > 
> > How did you measure this?
> > 
> 
> You could probably do it using conventional IP pings, I was doing it
> while testing an implementation of the Ethernet v2.0 Configuration
> Testing Protocol I wrote for the Linux kernel, which basically provides
> ethernet layer ping, so didn't have the overheads of IP packet
> processing or firewalling. I ran a tcpdump on the interface in question,
> and then compared the in and out timestamps. I think it was also
> measuring the effects of the CPU instruction case. IIRC, with a fast
> packet rate, the first in/out difference was around 40 microseconds,
> then would drop to between 4 and 8 microseconds. My guess at the time
> was that the fast packet rate was keeping the protocol code in the
> instruction cache. I didn't investigate further. Measuring in/out time
> differences at the source interface, subtracting in/out latency on the
> destination interface, indicated cable and/or switch latency.
> 
> > In all of this discussion (great btw) the actual software architecture
> > hasn't really been mentioned. IME that's usually where all the
> > bottlenecks come from, rarely the LAN...
> > 
> > 
> > -- 
> > 
> > 
> > Narelle
> > narellec at gmail.com
> _______________________________________________
> AusNOG mailing list
> AusNOG at lists.ausnog.net
> http://lists.ausnog.net/mailman/listinfo/ausnog