Hi Tom,

Comments inline below:

> The problem is that most Internet and subscriber traffic is using a
1500MTU or smaller frame.  So in theory, its would be just as efficient and
fast to bond two 100 mbps fiber connections than it would to buy 1- 1GB
fiber connection.

No, it wouldn't be.

> So the question is.... How do we most efficiently use 1GB fiber to get the
advantage of the full 1GB of capacity?

Simple: Install hardware that can pass 1 Gbps.

> Do we need to use some sort of packet agreegation/stuffing technology?


> Is GB etherner pointless for Internet transit backbones?


> Is GB just good for high capacity Transports, recognizing that routers
will likely split traffic to different smaller bandwidth peers?


> Is there a special router or router feature used to solve this problem?

No.  No special features or hardware are necessary.

> Is that method available to Linux?

Yes, though Linux is designed for SERVER performance, so a stock kernel will
likely also present some other bottlenecks.

> The reason I ask is several fold.  In a network design where all traffic
flows to a single source (for example many 100mbps baclhauls to remote areas
to 1 central data center), it would be beneficial  because the cost of 1 big
1GB pipe could be shared to deliver capacity to everything, better apt to
handle peak traffic and get higher oversubscription rates.   However, if teh
GB INternet pipe can not be efficiently used, this method would be severally
flawed. It might be better to have multipel 100mvps transit connections
spread out across one's network, so there was a shorter path to transit, and
the network's bandwdith spread out amungst multiple 100mbps transit
connection, for better over all throughput.  In other words, in a 10 city
network, 1- 100mbps pipe in each of teh 10 cities would allow a full
combined 1 gbps of Internet transit, where as agregating 100mbps from each
city to one central source where their was a single 1GB transit, would
result in only a 200mbps throughput, assuming traffic was delivered to it as
a 1500 MTU. 

> Any feedback?

> Take note that my comment that a 1500MTU frame 1 Gbps Ethernet card could
only pass 200kbps was based on some lab tests.

The solution is simply a hardware upgrade for starters.  A 32-bit/33 MHz bus
will top out around 200 Mbps.  If you look for a bus with higher speed
slots, you can triple your throughput without adjusting ANYTHING in your
Linux kernel.

If you want wirespeed GigE performance with multiple packet streams at a
more reasonable packet size (remember that 1500 byte packets aren't
realistic), you'll need to make some adjustments to the kernel.

> With the 1500MTU frame acheiving only 200kbps, our routers CPU utilization
was less than 20%, so it was not a saturated router.

Actually, it is.  Utilities top and vmstat don't necessarily reflect all of
the CPU utilization, and can't account for PCI bus contention/overhead.

> The second we changed MTU to 9600, we got over 800 mbps, and CPU
utilization was still very low, forget exact number but under 40%.

Well, sure...you've just taken your packet rate down significantly.  You
still have bus overhead and the limitations of a 32/33 bus (1 Gbps burst



Tom DeReggi
RapidDSL & Wireless, Inc
IntAirNet- Fixed Wireless Broadband 

WISPA Wireless List: wireless@wispa.org


Archives: http://lists.wispa.org/pipermail/wireless/

WISPA Wireless List: wireless@wispa.org


Archives: http://lists.wispa.org/pipermail/wireless/

Reply via email to