Re: [ofa-general] Re: IPoIB forwarding

[Rick Jones <rick.jones2@hp.com>]

Bryan Lawver wrote:
> I hit the IP NIC over the head with a hammer and turned off all offload 
> features and I no longer get the super jumbo packet and I have symmetric 
> performance.  This NIC supported "ethtool -K ethx tso/tx/rx/sg on/off" 
> and I am not sure at this time which one I needed to whack but all off 
> solved the problem.

Yeah, that does seem like a rather broad remedy, but I guess if it works... :) 
And I suppose most of those offloads don't matter for a NIC being used in a router.

Only problem is we don't know if it worked because it slowed-down the 10G side 
or because it had LRO disabling as a side-effect. If I were to guess, of those 
things listed, I'd guess that receive cko would have that as a side effect.

Just what sort of 10G NIC was this anyway?  With that knowledge we could 
probably narrow things down to a more specific modprobe setting, or maybe even 
an ethtool command, for some suitable revision of ethtool.

rick jones

> 
> Thanks for listening and re enforcing my search process.
> 
> bryan
> 
> At 01:32 PM 4/27/2007, Rick Jones wrote:
> 
>> Bryan Lawver wrote:
>>
>>> Your right about the ipoib module not combining packets (I believed 
>>> you without checking) but I did never the less.  The ipoib_start_xmit 
>>> routine is definitely handed a "double packet"  which means that the 
>>> IP NIC driver or the kernel is combining two packets into a single 
>>> super jumbo packet.  This issue is irrespective of the IP MTU setting 
>>> because I have set all interfaces to 9000k yet  ipoib accepts and 
>>> forwards this 17964 packet to the next IB node and onto the TCP stack 
>>> where it is never acknowledged.  This may not have come up in prior 
>>> testing because I am using some of the fastest IP NICs which have no 
>>> trouble keeping up with or exceeding the bandwidth of the IB side.  
>>> This issue arises exactly every 8 packets...(ring buffer overrun??)
>>> I will be at Sonoma for the next few days as many on this list will be.
>>
>>
>>
>> Some NICs (esp 10G) support large receive offload - they coalesce TCP 
>> segments from the wire/fiber into larger ones they pass up the stack.  
>> Perhaps that is happening here?
>>
>> I'm going to go out a bit on a limb, cross the streams, and include 
>> netdev, because I suspect that if a system is acting as an IP router, 
>> one doesn't want large receive offload enabled.  That may need some 
>> discussion in netdev - it may then require some changes to default 
>> settings or some documentation enhancements.  That or I'll learn that 
>> the stack is already dealing with the issue...
>>
>> rick jones
>>
>>> bryan
>>>
>>> At 11:06 AM 4/26/2007, Michael S. Tsirkin wrote:
>>>
>>>> > Quoting Bryan Lawver <lawver1@llnl.gov>:
>>>> > Subject: Re: IPoIB forwarding
>>>> >
>>>> > Here's a tcpdump of the same sequence.  The TCP MSS is 8960 and it 
>>>> appears
>>>> > that two payloads are queued at ipoib which combines them into a 
>>>> single
>>>> > 17920 payload with assumingly correct IP header (40) and IB header
>>>> > (4).  The application or TCP stack does not acknowledge this 
>>>> double packet
>>>> > ie. it does not ACK until each of the 8960 packets are resent
>>>> > individually.  Being an IB newbie, I am guessing this combining is
>>>> > allowable but may violate TCP protocol.
>>>>
>>>> IPoIB does nothing like this - it's just a network device so
>>>> it sends all packets out as is.
>>>>
>>>> -- 
>>>> MST
>>>
>>>
>>> _______________________________________________
>>> general mailing list
>>> general@lists.openfabrics.org
>>> http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
>>> To unsubscribe, please visit 
>>> http://openib.org/mailman/listinfo/openib-general