Matt and I have had problems with BEE2 - switch with short cables. The lanes lose sync, and then take a fixed time to resync. Data is lost during this process, hence the losses are higher at higher datarates. With lower rates, it is able to buffer the data and resume without significant losses.

FWIW, with default PHY settings (max preemph and swing), my BEE2 cable length preferences are as follows:
BEE2 - IBOB: 1m
BEE2 - switch: 3m
BEE2 - Myricom NIC: anything from 1m to 3m.

I believe we can improve the range of cable lengths that work, but you'll have to tweak the PHY's preemph and swing settings. Unfortunately, on the V2Ps, this must be done at compile time. ROACH lets you do this at runtime, which aids tuning.

I would be surprised to find that the switch is causing the problem. I am currently pumping 8Gbps on each of 8 ports on my XG2000C without a single dropped packet.

Jason

On 25 Nov 2009, at 02:00, Paul Demorest wrote:

Hey guys,

Jumbo frames are already enabled on the switch. The packet loss we're seeing is about 0.1%, so most of the data is making it through. The weird thing is that this number seems independent of the data rate. Even at <1 MB/s we still lose 0.1% of the data.

As John already said, connecting the bee2 directly to the receiving computer results in zero packet loss at low data rates. We do lose some if the data rate is set much higher (>~400 MB/s depending on the data processing, system load, etc). But that kind of thing is expected..

-Paul

On Tue, 24 Nov 2009, Peter McMahon wrote:

Page 3 of Jason's switch memo at
http://casper.berkeley.edu/memos/switch_configuration.pdf has the commands
to enable jumbo frame on some of the Fujitsu switches.

Peter

-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Dan Werthimer
Sent: Tuesday, November 24, 2009 3:34 PM
To: John Ford
Cc: [email protected]
Subject: Re: [casper] 10 GBe Switch weirdness


hi john,

i vaguely remember that there's a command you need to issue
to the fujitsu switches to enable jumbo packets ??

jason will know.

dan



On 11/24/2009 02:01 PM, John Ford wrote:
Hi all.  We've been running GUPPI for some time now with a direct
connection from our bee2, "bee2" to our host, "beef". Works fine, no
dropped packets, etc.  Life's fine.

To build our next machine, GUPPI-2, we decided to insert a Fujitsu
XG-2000C switch between them, and now we are losing packets.  Same
hardware, same software, except the switch is now between the bee2 and the host. We have very short cables on all the ports. 1 meter or 0.5 meter. Changing cables had no effect. Changing switch ports had no effect. The ports are set to use jumbo packets. Moving the cable from the switch back
to the host fixes the problems, and it all magically works again.

Other hosts on the 10 gbe network do not seem to suffer from packet loss. The switch's monitoring screens do not show any packet loss in the switch.

Does the switch do anything to the timing of the packets? Any ideas about
incompatible NIC's?  They are all myricom pcie cards.

Any other ideas?

John









Reply via email to