[ovs-dev] [PATCH v3] lib/dpif-netdev: Integrate megaflow classifier.

[Simon Horman]

On Tue, Oct 28, 2014 at 02:47:00PM -0700, Alex Wang wrote:
> Hey Simon,
> 
> I did the following test:
> 
> - Environment:
> 
>    host 1: 7 pmd threads, each pmd thread handling one rx queue from
>    interface dpdk1
> 
> 
>    bash-4.3# ovs-appctl dpctl/show
>    netdev at ovs-netdev:
>            lookups: hit:778832 missed:3242727 lost:0
>            flows: 89270
>            port 1: br0 (tap)
>            port 3: dpdk1 (dpdk: configured_rx_queues=7,
>                                   configured_tx_queues=16)
>            port 4: int_port (tap)
>            port 0: ovs-netdev (internal)
> 
> 
>        Bridge "br0"
>            fail_mode: standalone
>            Port "dpdk1"
>                Interface "dpdk1"
>                    type: dpdk
>            Port "br0"
>                Interface "br0"
>                    type: internal
>            Port int_port
>                Interface int_port
>                    type: internal
>        ovs_version: "2.3.90"
> 
> 
>    10 netperf tcp_crr sessions from dpdk1 on host 2 (dpdk1 iface are
>     directly connected between host 1/2)
> 
>     flow limit is set to 200000,
> 
>     so in this test, the contention is between pmd thread trying to install
> flows
>     to classifier/flow_table, and revalidators trying delete flows from the
>     the same structs...
> 
> - Result:
> 
> *   with this patch:*
>    175 Trans/s * 10 (TCP_CRR sessions) (without megaflows)
> 
>     bash-4.3# ovs-appctl upcall/show
>     netdev at ovs-netdev:
>             flows         : (current 145961) (avg 147542) (max 153922)
> (limit
>                                 200000)
>             dump duration : 270ms
> 
> 
> *   without this patch:*
>    110 Trans/s * 10 (TCP_CRR sessions) (without megaflows)
> 
>     bash-4.3# ovs-appctl upcall/show
>        netdev at ovs-netdev:
>            flows         : (current 89296) (avg 91497) (max 107460) (limit
> 200000)
>            dump duration : 184ms
> 
> So, since this patch reduce the overhead of classifier operations, the
> balance point between flow insert rate/delete rate goes up,

Hi Alex,

thanks for those numbers. It does seemed to be quite an improvement.