On Thu, Sep 18, 2008 at 12:03 PM, Fab Tillier
<[EMAIL PROTECTED]> wrote:
>>> I understood that Fab checked this issue (by 10 retries of 1 second TO)
>>> and found that it didn't help there. Yet another try can be enlarging
>>> the TO to be 5 sec and sending less retries
>>
>> I think some exponential backoff strategy with some randomization
>> might be better.
>
> The problem with this is that the layers above IPoIB (namely the network 
> stack generating ARP requests and expecting ARP responses) doesn't have 
> visibility into this backoff strategy, and will give up on an ARP request if 
> the response doesn't come back in time.  The response could be delayed for a 
> long time if the SM isn't responding to queries in a timely manner, since 
> IPoIB needs to resolve the path in order to send the unicast response.  I 
> don't know the timeout for an ARP response, but I'd be surprised if it was 10 
> seconds, let alone whatever you would get with exponential backoff.
>
> I initially tried exponential backoff to resolve the problem I was seeing 
> with these MPI apps, and it didn't work because of this.  That's when I set 
> out on a path to take the SM out of the equation as much as possible.

Yes, I think we went through this before :-(

10 seconds is a long time to wait for the SM to respond anyhow.

I understand why ofw is bypassing SA but in the long term this is a
_bad_ thing IMHO. I think the real solution needs to be in increasing
the SA scalability. I don't think there's been any serious effort on
that. I'd also be concerned about whether all SM flavors in some
configuration need to support a certain sustained (and peak)
transaction rates.

-- Hal

>
> -Fab
>
_______________________________________________
ofw mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ofw

Reply via email to