XAPP1082 v3.0 2014.4

This wiki page summarizes the performance of PS-EMIO and PL Ethernet (with/without) CSO and jumbo frame support.
Build:
Vivado 2014.4
Kernel version 2014.4 (3.17)

These measurements are obtained against Fedora-20 high performance peer machine. The NIC on motherboard has been used with default offload options
(GSO, TSO) enabled. Performance is expected to differ when a different peer OS is used and peer NIC offload options are disabled.
Socket size in netperf (-s) is configured to 65536 for inbound tests.

XAPP1082_2014.4_PS_EMIO_perf.png
Figure-1 : PS GEM Performance
xapp1082_pl_eth_2014_4.png

Figure-2 : PL Ethernet Performance
Note: Coalescing of packets may be observed even with '-D' option in netperf. To observe appropriate performance with smaller packets, one can also restrict MTU to smaller values. For MTU calculation include TCP/IP and MAC header in addition to message size. For example for 64B message size, set MTU to 130 (=14+20+32+64). The following commands can be used to set different MTU values-
ifconfig eth1 down
ifconfig eth1 mtu 130
ifconfig eth1 up


XAPP1082_2014.4_PL_ethCSO.png
Figure-3 : PL Ethernet CSO impact on Throughput & CPU Utilization

XAPP1082_2014.4_PL_Eth_Jumbo.png
Figure-4 : PL Ethernet Jumbo Frame Performance