On Thu, Oct 5, 2017 at 5:46 PM, Andrew Lunn <and...@lunn.ch> wrote: > On Thu, Oct 05, 2017 at 04:23:45PM -0400, Patrick Talbert wrote: >> Network performance can suffer when a load balancing bond uses slave >> interfaces which are in different NUMA domains. >> >> This compares the NUMA domain of a newly enslaved interface against any >> existing enslaved interfaces and prints a warning if they do not match. > > Hi Patrick > > Is there a bonding mode which might actually want to do this? Send on > the local domain, unless it is overloaded, in which case send it to > the other domain? >
I suppose there could theoretically be a bonding mode that could do that, but currently no such mode exists. > There is also this talk for netdev: > > https://netdevconf.org/2.2/session.html?shochat-devicemgmt-talk >From reading the abstract there, it sounds like such a device driver would want to abstract away the numa location of the underlying devices from the "unified" net device it presents to the kernel. > > Andrew My goal with the patch is not to prevent some one from bonding whichever interfaces they want, only to notify them that what they are doing is *likely* to be less than ideal from a performance perspective. Even if some theoretical load balancing bonding mode was intelligent enough to consider NUMA when choosing a transmit interface, it never has control over the interface traffic is received on (excluding the strange balance-alb mode). Patrick
