[ovs-dev] CPU performance measurement - OvS+DPDK

[Ilya Maximets]

On 19.09.2018 06:42, Shahaji Bhosle wrote:
> Hi Ilya,
> Just trying to figure out how much the CPU processing cycles are being used using the "dpif-netdev/pmd-stats-show" command and what I am seeing that when I start the test idle is 100% and then processing cycles just keep increasing as long as I keep running the test, I was hoping it to stabilize at some point but it does not. Is there any other way to find out how much CPU is being utilized for the test.

Are you running the "dpif-netdev/pmd-stats-clear" before the test?
Maybe you system just was idle for a long time.
My personal suggestion is to run it few seconds after starting the
test to have more accurate short term numbers.

> 
> Also, I am seeing drop and errs in PHY-to-PHY, coming ixia --dpdk128<-ovs+dpdk-> dpdk129 -- ixia, just trying to understand whats going on with this. I do have about 8K IP lookups, the actions is go to the other port with mod_dl_dst of the next_hop.

Are you changing the source address too? It's not important if packet
generator is connected directly, but could be an issue if you have
some L2 learning switch between test machine and the generator.

> 
> OFPST_PORT reply (xid=0x2): 3 ports
>   port 129: rx pkts=28188204935, bytes=3337048521914, drop=0, errs=5381086, frame=?, over=?, crc=? 
>            tx pkts=28251930544, bytes=3340826046788, drop=*52764491*, errs=0, coll=?
>   port LOCAL: rx pkts=0, bytes=0, drop=0, errs=0, frame=0, over=0, crc=0
>            tx pkts=0, bytes=0, drop=0, errs=0, coll=0
>   port 128: rx pkts=28305064186, bytes=3344059379368, drop=0, errs=*21967104*, frame=?, over=?, crc=?
>            tx pkts=28188542419, bytes=3337022565678, drop=*32176*, errs=0, coll=?

You may use following command to better understand the Rx/Tx errors:
# ovs-vsctl get interface ens1f0 statistics

> Thanks, Shahaji
>