Intel IXP12xx manual Hardware 40-byte packet performance, 138 142 144, 9.5, 88,300

Version 1.0, 4/10/02

degrade in these scenarios, and the design becomes subject to ATM overflows from running “_VolgaGetChanCounters”.

Ethernet	ATM	IXF6012	ATM	IXF6012
Input	Transmit	Transmit	Receive	Overflows
Ports	Rate [%]	Idle	Ports
8	84	N/A	1	0
7	73	N/A	1	0
6	63	N/A	1	0
0	0	N/A	1	0

Figure 6 – Single-cell/PDU Performance using 143MHz DRAM

Ethernet

Transmit

KFrame/s

138 - 147

142 - 148

144 – 148

148,808

Ethernet

Transmit

[MB/s]

8.8– 9.4

9.0– 9.5

9.2– 9.5 9.5

Repeating the same measurements for 143 MHz DRAM results in the same ATM transmit bandwidth in all cases (Figure 6). Interference from “_VolgaGetChanCounters” is gone however, even in the 8 port configuration. Ethernet Transmit performance is slightly better, but still sub- wire-rate in the 8 port configuration. When there is no Ethernet input at all, the ratio between packets dropped by the IP Router versus full Ethernet transmit queues improved such that the Ethernet Transmit queues drop 45% and the IP Router drops only 55% of the excess input.

Hardware 40-byte packet performance

Ethernet	ATM	IXF6012	ATM	IXF6012
Input	Transmit	Transmit	Receive	Overflows
Ports	Rate [%]	Idle	Ports
8	97	30,000	1	700
7	100	1,500	1	69
6	100	700	1	0
5	100	0	1	0
4	84	N/A	1	0

Figure 7 – Two-cell PDU Performance on 133MHZ DRAM

Ethernet

Transmit

KFrame/s

88,300

88,300

88,300

88,300

88,300

Ethernet

Transmit

[MB/s]

5.6

5.6

5.6

5.6

5.6

For the 40-byte (2 cell/PDU) workload there are half as many IP lookups/second as are required in the 29-byte (1 cell/PDU) workload. As expected, the IP Router was able to keep up with this workload and didn’t drop any packets.

For five Ethernet input ports, the design performed perfectly, and “_VolgaGetChanCounters” did not cause any dropped cells.

Reducing the Ethernet input to 4 ports did not allow enough input to saturate ATM Transmit. Increasing the Ethernet input ports to 6, 7 and 8 allowed over-subscription in the Ethernet to ATM forwarding direction, as evidenced by the ATM Transmit idle cells and ATM Receive overflows. Indeed, even under nominal conditions, the 8 port input configuration was able to drive only 97% of the ATM wire.

For this packet size, Ethernet is more efficient that ATM, and the Ethernet transmitter cannot be expected to be driven at wire rate. Indeed, under nominal conditions the Ethernet transmitter performed perfectly for all tested configurations.

Page 11 of 17