[Atlas-anchors-pilot] Heads-up about new HW specs

[Nick Hilliard]

On 29/08/2013 15:59, Vesna Manojlovic wrote:
> we are currently testing:
> - Soekris box net6501-70 Board
> - in 1U 19 Inch Rackmount Case
> - and 128 GB SSD SLC

we have a couple of 6501-30 units in production at INEX.  They are very
good at what they do, and the power draw is a major attraction for us.  But
before the RIPE NCC makes a commitment to using them, I would recommend
benchmarking them very carefully with either packet generators or else mgen
(http://downloads.pf.itd.nrl.navy.mil/mgen/) to make sure that you
understand their throughput and pps limitations.  I haven't benchmarked the
6501 units yet, but the previous generation Soekris 5501 units start
dropping packets at ~10 mbit/sec or 4000 pps, which ever comes first.  The
6501 units will behave much better than this, but I don't have figures for
them.

Also, the default ethernet cards on these boxes are based on the Intel
82574 chipset.  There are some really nasty bugs on this chipset and even
on a modern bus on a big machine, they will start dropping packets at
around 400 mbit/sec in their default configuration.  You can work around
with some of this brokenness by disabling active state power management.
On linux this can be done using the following command:

> bash# for i in `lspci | grep 82574 | awk '{print $1}'`; do setpci -s ${i} CAP_EXP+10.b=40; done

Dunno what the freebsd equivalent is, but I'm sure it's as obvious as the
linux fix.  I haven't got them to reliable handle iperf >= 650mbit/sec on
any configuration, but obviously ripe anchor nodes won't need to do this
amount of traffic so it won't be an issue.

The testing I've done on 82574 based ethernet cards was not done on soekris
kit, btw.

Also you would want to confirm that these cards don't suffer from:

http://blog.krisk.org/2013/02/packets-of-death.html

Nick