Wednesday, January 10, 2007

[en] Throughput tests WRTSL54GS v1.1 vs WRT54GS v3.0

Throughput tests WRTSL54GS v1.1 vs WRT54GS v3.0

The goal of this test is to measure and compare the NAT'ing/routing/switching behavior of these two WRT54G's [1] running OpenWRT Whiterussian RC6 [5]:

models/nCPUflash/RAMeth switch
WRTSL54GS v1.1 [2]CJK11xxxxxxxBroadcomm 4704rev9 266Mhz8/32MBinCPU
WRT54GS v3.0 [3]CGN5xxxxxxxBroadcomm 5352 200Mhz8/32MBBCM5325

IMPORTANT: nbd ( openwrt developer) pointed a problem with "entropy gathering which fills /dev/random from wifi" in WR rc6 and suggested me trying a newer build (09test); I'll try it an repost the results (specially the wifi ones)
--jjo; Jan 10, 2007.

Testbed, methodology

  • Topology

  •                        --------> iperf test connection------>
    WIFI-NATed [CLIENT]+)) wifi (( +[ WRT ]+=== wan ===+[SERVER] (iptables MASQUERADE'd)
    LAN-NATed [CLIENT]+=== lan ===+[ WRT ]+=== wan ===+[SERVER] (iptables MASQUERADE'd)
    LAN-routed [CLIENT]+=== lan ===+[ WRT ]+=== wan ===+[SERVER] (plain routing)
    LAN-switched [CLIENT]+=== lan ===+[ WRT ]+=== lan ===+[SERVER] (hardware switching)

    "lan": any LAN (1of4) port
    "wan": WAN port
    "wifi": 802.11g link at 54Mbps (aprox 2mts distance between antennas)
    iptables MASQUERADE'd: default whiterussian setup (masquerade both wifi,lan to wan)
    plain routing: accomplished with iptables full "flushing" + server route to client via WRT

  • Hardware

    • client: IBM Thinkpad T42, lan: e1000 (10/100/1000), wifi: ipw2200
    • server: generic PC sempron 2200+, lan: 8139too
    • routers: (stated above)

  • Software
    • client, server: iperf [4] (-c, -s) on linux-2.6, glibc-2.3+.
    • routers: OpenWRT WhiteRussian RC6 [5].


Throughput values in megabits per second taken from iperf's output. Please note that measured rates by iperf correspond to _application_, ie: they don't consider protocol headers, etc.

CPU usage % shown by top for ksoft_IRQd from a ssh logged session, this "top" session was not considered for throughput values (separately taken) to avoid "measuring load influence": 100% CPU usage was aprox= 90% ksoft_IRQd + 8% dropbear + 2% top.

Throughput (CPU usage %)
WRTSL54GS23.5mbit/s (65%)47.5mbit/s (90%+)55.8mbit/s (90%+)93.7mbit/s (0%)
WRT54GS 10.5mbit/s (14%)34.2mbit/s (90%+)40.4mbit/s (90%+)93.8mbit/s (0%)



WIFI-NATed at 23.5 mbits (802.11g @54Mbps) with ~65% CPU usage denotes "reasonable" CPU availability for this 802.11g saturated link.

LAN-NATed at 47.5 mbits shows the CPU bottleneck when routing+MASQUERADING, considering that ip_conntrack table has just a few entries this is a VERY optimistic max.

LAN-routed at 55.8 mbits is pretty poor, I think that ip_conntrack processing is ALSO affecting this scenario (if loaded it gets called even for non-nated streams); it would be veryVERY useful to have ip_conntrack optionally loaded by openwrt (currently static).

LAN-switched at 93.8 mbits is expectable given that hardware only switching is involved.


LAN-NATed and LAN-routed values give about 75% when compared to its newer cousin, 75% is roughly 200/266, respective CPU clocks [MHz] (this confirms CPU saturation).

WIFI-NATed at 10.5 mbits (dbl-confirmed 802.11g @54Mbps) is a bad smelling frog, I double checked openwrt setup... nothing strange found; about 1mo ago I was running RC5 getting ~20mbits+ rates (informal, from rsync/ftp rates), so it maybe something with RC6 wifi in this hardware, dunno(?).

As expected, LAN-switched gave same results.


Both WRT's seem reasonable for 802.11g @54Mbps even when NAT'ing traffic (common usage), obviously depending on traffic patterns, how many clients, etc ... given the constrained RAM (and CPU) for NAT+conntracking ... NOT tested here.

On the contrast, their CPU is not powerful enough to saturate _both_ (LAN+WAN) 100mbps links simultaneously, even when "only" routing (55mbit/s max on WRTSL54GS); this could be enhanced by having ip_conntrack optionally (not)loaded by openwrt.

Out of band disclaimer

Ok ... I know, you agree: blogger sucks for writing technical reports .. sheesshhhh :-P


[4]Iperf bandwidth measurement tool:

This thread at openwrt forums talks about openwrt experiences on Linksys WRTSL54GS model, and was the starting point to encourage loading openwrt on WRTSL54GS v1.1 model.