We have a small network of routers and switches between 3 distinct sites. The Cisco devices consist of 1841, and 2621 routers; and 2950 switches.
Interconnecting the sites are 2 telco LAN Extension services rated at 100 Mbit/s. From end to end is 100Mbit interfaces / technologies.
It has been noticed that there is a consistant drop in throughput depending on what network device is traversed.There is no other traffic present yet.
e.g. An 1841 router with 2 fastethernet ports results in a throughput of 85Mbits/s from 1 subnet to another. The 2 network devices were baselined on the same network segment at 92Mbits/s which shows a 7 Mbits/s drop throgh a router with basic configuration,no routing protocols running, and no other traffic flowing. The links between the Cisco router and 2950 switches are configured as 100 and full duplex (no errors showing).
I have similar drops in throughput when 2950 switches are connected to each end of the telco LAN Extension where there is a 8Mbits/s drop in throughput.
The most dramatic drop in throughput is if a 2621 router is connected to the switches on each end of the LAN Extension.
The resulting throughput drops by 34Mbits/s. Again router has basic config, EIGRP running for 1 test and replaced by static routes for retest with no real deviation of throughput drop.
Am I over optimistic to think the throughputs should be close to the maximum throughput they are designed for, especially when they not subjected to normal user traffic ?
The performance tool used is netperf and seems to give reasonable accurate results.
The only surprising result is the 2950 one that is supposed to be able to switch L2 at wire rate.
The routers, 1841 and 2621 especially that is quite old and weak, have never been advertised nor designed for 100 mpbs wire speed throughput. You need to scale up of a class or two for that.
In particular your 1841 result is quite close to what Miercom has independently tested - 130 Mbps for bidirectional, 1460 Btyes/packet with NAT and Fw enabled.
Hope this helps, please remember to rate all useful posts
Thanks for your reply.
The 1841 is a relatively new router aimed at small and medium businesses, please see present Cisco description on website:
The award-winning Cisco 1800 Series integrated services routers, ideal for small to medium-sized businesses and small enterprise branch offices, enable businesses to reduce costs by deploying a single, resilient system for fast, highly secure, delivery of multiple mission-critical business services.
The Cisco 2621 router is actually a 2621XM which is a more recent edition and more powerful.
The network we have, at best, would be small/medium sized network.
The ability to transfer data close to 100Mbits/s shouldn't be beyond the ability of these devices especially when under no load whatsoever.
A possible upgrade is to a Layer 3 switch (probably 3560-E). Do you know if this will provide 100Mbits/s throughput? If not, what would you suggest?
Do you have any advice on what I could try with the 2950 issue with throughput loss?
What are the machines that you are testing as source and destination? The reason I ask is you cannot get 100mbit/s out of a 100mbit network card.
Another thing to keep in mind is RTT affects throughput, as well as, many other factors.
You are quite right in what you say, that you can't get 100Mbits/s from a 100Mbit network card. The 2 test machines were initially dirctly connected to each using a crossover cable and netperf was used to gauge their throughput. They did so at approx 92Mbits/s.
This value was used as the baseline to measure any other tests up against.
I believe netperf gives a reliable reflection on throughput capabilities of devices.
since you are benchmarking throughput... test some other tools as well to see their results and it will give you some more datapoints.
iperf - http://picasaweb.google.com/home
bwctl (which uses iperf) - http://e2epi.internet2.edu/bwctl/
That being said, when connected via xover cables the latency is virtually nothing... test the latency when you connect through the router and switches to see what that metric is, because latency effects tcp and throughput. Also test udp v/s tcp to see the differences.
I'd be interested to see, because I do the same throughput testing but with gigabit and 10 gigabit these days ;)
1841 - we discussed already.
2621XM - I can categorically exclude that you will be able to squeeze more performance from it. Too slow CPU and bus bandwidth limitations too.
2950 - suggest you upgrade to latest image and try again - check queues and QoS settings, these sometime can cause problems.
A 3560 (now available in a new 8 ports desktop model) would certainly be able to handle FE and GE at wire speed in full routing mode.