On 12/04/2013 07:28 AM, Ben Pfaff wrote:
I'm surprised you're not CCing Thomas then.
Thomas's setup and results look pretty clear to me. pktgen to generate
traffic. Without the patch, 74 kpps, with the patch 206 kpps, roughly
3x performance.
On Wed, Dec 04, 2013 at 06:22:17AM +0000, Zhuangyuxin wrote:
Actually, the result I mentioned is from Thomas' review. Here's the
whole snippet. I'm wondering how to test and measure the unicast
netlink I/O performance improvement.
I'm glad to share additional details.
Attached is the script to configure and run pktgen. You have to modify
CPU_MAX and CPU_USE to match the number of cores and number of threads.
You want to leave some cores to run upcall threads as well.
The script as-is produces jumbo frames and no checksumming is performed
as it would happen for offloaded checksummin. Setting the flag "udpsum"
will triger UDP checksum calculation in the upcall.
OVS setup:
ovs-vsctl add-br br0
ovs-vsctl add-port br0 int0 -- set Interface int0 type=internal
#!/bin/bash
modprobe pktgen
function pgset() {
local result
echo $1 > $PGDEV
}
CPU_MAX=7
CPU_USE=4
OUTDEV="int0"
# Cleanup
for processor in $(seq 0 $CPU_MAX)
do
PGDEV=/proc/net/pktgen/kpktgend_$processor
pgset "rem_device_all"
done
for ((processor=0;processor<$CPU_USE;processor++))
do
PGDEV=/proc/net/pktgen/kpktgend_$processor
pgset "add_device $OUTDEV@$processor"
PGDEV=/proc/net/pktgen/$OUTDEV@$processor
pgset "count 5000000"
pgset "flag QUEUE_MAP_CPU"
pgset "pkt_size 9000"
pgset "dst 10.0.0.1"
pgset "dst_mac 5e:ed:0f:2c:cb:40"
pgset "flag IPDST_RND"
pgset "dst_min 10.0.0.0"
pgset "dst_max 10.255.255.255"
pgset "flows 5000000"
pgset "flowlen 1"
done
# Time to run
PGDEV=/proc/net/pktgen/pgctrl
echo "Running... ctrl^C to stop"
pgset "start"
echo "Done"
for ((processor=0;processor<$CPU_USE;processor++))
do
cat /proc/net/pktgen/$OUTDEV@$processor
done
_______________________________________________
discuss mailing list
discuss@openvswitch.org
http://openvswitch.org/mailman/listinfo/discuss
