[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH] x86/apic/x2apic: Share IRQ vector between cluster members only when no cpumask is specified

>>> Andrew Cooper <andrew.cooper3@xxxxxxxxxx> 08/07/17 6:21 PM >>>
>On 07/08/17 17:06, Jan Beulich wrote:
>>>>> Andrew Cooper <andrew.cooper3@xxxxxxxxxx> 08/07/17 5:40 PM >>>
>>> On 07/08/17 16:38, Jan Beulich wrote:
>>>> So a single NIC drove the system out of vectors? That's insane, I would
>>>> say, i.e. I'd call this a misconfigured system. But yeah, if we really want
>>>> to get such a thing to work despite the insanity ...
>>> No.  That's one single (dual headed) card with 128 virtual functions each.
>> But putting such in a single-core system is, well, not very reasonable. At
>> the very least I'd expect the admin to limit the number of VFs then, or not
>> load a driver in Dom0 for them (but only for the PF).
>Its not a single core system, but cluster mode allocates vectors across
>clusters, not cores. This turns even the largest server system into a
>single core system as far as vector availability goes.
>There was definitely a change in behaviour between Xen 4.1 and Xen 4.4. 
>Looking through the ticket, it appears the regression was introduced by
>your cluster IPI broadcast optimisation (which iirc was for the 64 vcpu
>windows issue), but I'm struggling to locate it in Xens history.

But that had nothing to do with the number of vectors used/needed.


Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.