[lvs-users] LVS and receive packet steering

Simon Horman horms at verge.net.au
Wed Jan 13 23:31:55 GMT 2010


On Wed, Jan 13, 2010 at 01:42:00PM +0200, Siim Põder wrote:
> Hi
> 
> Currently LVS scales to the maximum of different network interfaces
> used for moving traffic. For DR it would most likely be 1 cpu core,
> for NAT 2 cores. So multicore machines can't really be used
> effectively (most cores will be idle).
> 
> I saw "receive packet steering" patch being mentioned a few weeks back
> (http://lwn.net/Articles/361440/). Essentially it hashes src/dst
> ip/port (for tcp and udp) to select the CPU that will handle any
> packet (hence keeping a flow on a single CPU core). Would also allow
> ipvs module code to be run on multiple processors in parallel (of
> course, it CAN run in parallel, but usually it will just not do it
> currently, as the incoming packets from an interface will be handled
> sequentially)?
> 
> I'd imagine that the patch should just transparently make general
> networking, ipvs (and netfilter) code scale to all CPU cores, unless
> I'm misunderstanding something about this patch. I was wondering if
> someone had already tried this or could speculate whether this should
> be the case?

Hi Siim,

that is my understanding too. But I also haven't had a chance to test it out.






More information about the lvs-users mailing list