On Fri, 21 Jan 2011, juice wrote:

I am using now a PCIe Intel e1000e card, that should be able to handle
the needed traffic amount.

The statistics that I get are as follows:

kernel 2.6.32-27 (ubuntu 10.10 default)
    pktgen:           750064pps 360Mb/sec (360030720bps)
    AX4000 analyser:  Total bitrate:             383.879 MBits/s
                      Bandwidth:                 38.39% GE
                      Average packet intereval:  1.33 us

kernel 2.6.37 (latest stable from kernel.org)
    pktgen:           786848pps 377Mb/sec (377687040bps)
    AX4000 analyser:  Total bitrate:             402.904 MBits/s
                      Bandwidth:                 40.29% GE
                      Average packet intereval:  1.27 us

kernel 2.6.38-rc1 (latest from kernel.org)
    pktgen:           795297pps 381Mb/sec (381742560bps)
    AX4000 analyser:  Total bitrate:             407.117 MBits/s
                      Bandwidth:                 40.72% GE
                      Average packet intereval:  1.26 us

your computation of Bandwidth (as Ben Greear said) is not accounting for 
the interframe gaps.  Maybe more useful is to note that wire speed 64 byte 
packets is 1.44 Million packets per second.

In every case I have set the IRQ affinity of eth1 to CPU0 and started
the test running in kpktgend_0.

The complete data of my measurements follows in the end of this post.

It looks like the small packet sending effiency of the ethernet driver
is improving all the time, albeit quite slowly.

Now, I would be intrested in knowing whether it is indeed possible to
increase the sending rate near full 1GE capacity with the current
ethernet card I am using or do I have here a hardware limitation here?

I recall hearing that there are some enhanced versions of the e1000
network card, such that have been geared towards higher performance
at the expense of some functionality or general system effiency.
Can anybody point me how to do that?

I think you need different hardware (again) as you have saddled yourself 
with a x1 PCIe connected adapter.  This adapter is not well suited to 
small packet traffic because the sheer amount of transactions is effected 
by the added latency due to the x1 connector (vs our dual port 1GbE 
adapters with a x4 connector)

As I stated before, quoting myself:

quoted

Which do you suppose is the reason for poor performance on my setup,
is it lack of multiqueue HW in the GE NIC's I am using or is it lack
of multiqueue support in the kernel (2.6.32) that I am using?

Is multiqueue really necessary to achieve the full 1GE saturation, or
is it only needed on 10GE NIC's?

with Core i3/5/7 or newer cpus you should be able to saturate a 1Gb link 
with a single core/queue.  With Core2 era processors you may have some 
difficulty, with anything older than that you won't make it. :-)

quoted

As I understand multiqueue is useful only if there are lots of CPU cores
to run, each handling one queue.

The application I am thinking of, preloading a packet sequence into
kernel from userland application and then starting to send from buffer
propably does not benefit so much from many cores, it would be enough
that one CPU would handle the sending and other core(s) would handle
other tasks.

Yours, Jussi Ohenoja


*** Measurement details follows ***


root@d8labralinux:/var/home/juice# lspci -vvv -s 04:00.0
04:00.0 Ethernet controller: Intel Corporation 82572EI Gigabit Ethernet
Controller (Copper) (rev 06)

My suggestion is to get one of the igb based adapters, 82576, or 82580 
based that run the igb driver.

If you can't get a hold of those you should be able to easily get 1.1M pps 
from an 82571 adapter.

you may also want to try reducing the tx descriptor ring count to 128 
using ethtool, and change the ethtool -C rx-usecs 20 setting, try 
20,30,40,50,60

Jesse