Cisco 3620 LAN routing performance Versus Windows 2000 Server

We recently purchased a Cisco 3620 to replace a Windows 2000 Server.
The only module that is loaded is a Dual Fast Ethernet NM-2FE2W.
This is being used as a LAN router (servers one side -- workstations
other).

Now, I have run a number of benchmarks to check the throughput of the
interfaces and interestingly have come up with the following:

Windows 2000 Server
-------------------

Diskbench, one thread, block size 15MB, blocks 10

Created file: \\192.168.1.5\temp\DiskBench.bin
Size: 157286400 bytes
Time: 118780 ms
Transfer Rate: 1.263 MB/s

Diskbench, one thread, block size 2MB, blocks 5

Created file: \\192.168.1.5\temp\DiskBench.bin
Size: 10485760 bytes
Time: 4846 ms
Transfer Rate: 2.064 MB/s

Create File Bench ended

Cisco 3620 Router - Full duplex 100Mbit
-----------------

Diskbench, one thread, block size 15MB, blocks 10

Created file: \\192.168.1.5\temp\DiskBench.bin
Size: 157286400 bytes
Time: 633501 ms
Transfer Rate: 0.237 MB/s

Diskbench, one thread, block size 2MB, blocks 5

Created file: \\192.168.1.5\temp\DiskBench.bin
Size: 10485760 bytes
Time: 33238 ms
Transfer Rate: 0.301 MB/s

So basically - a very big difference in throughput.
100Mbit/s divided by 8 bits is 12MB/sec.
Take into account 20% overhead for TCP and we are down to 9.6MB/sec.
Full duplex should alloow 9.6MB/sec either way.

I then did a test by forcing the interfaces into half duplex mode:

Diskbench, one thread, block size 2MB, blocks 5

Created file: \\192.168.1.5\temp\DiskBench.bin
Size: 10485760 bytes
Time: 7540 ms
Transfer Rate: 1.326 MB/s

Still below the Windows 2000 Server routing performance, but improved
from full duplex mode.

Output from show ver:

Cisco Internetwork Operating System Software
IOS (tm) 3600 Software (C3620-I-M), Version 12.1(5)T7, RELEASE
SOFTWARE (fc1)
TAC Support: http://www.cisco.com/cgi-bin/ibld/view.pl?i=support
Copyright (c) 1986-2001 by cisco Systems, Inc.
Compiled Mon 16-Apr-01 21:43 by ccai
Image text-base: 0x60008950, data-base: 0x60A28000

ROM: System Bootstrap, Version 11.1(20)AA2, EARLY DEPLOYMENT RELEASE
SOFTWARE (f
c1)
ROM: 3600 Software (C3620-I-M), Version 12.1(5)T7, RELEASE SOFTWARE
(fc1)

C3620-ROUTER1 uptime is 2 weeks, 1 day, 9 hours, 30 minutes
System returned to ROM by power-on
System image file is "flash:c3620-i-mz.121-5.T7.bin"

cisco 3620 (R4700) processor (revision 0x81) with 27648K/5120K bytes
of memory.
Processor board ID 13891241
R4700 CPU at 80Mhz, Implementation 33, Rev 1.0
Bridging software.
X.25 software, Version 3.0.0.
2 FastEthernet/IEEE 802.3 interface(s)
DRAM configuration is 32 bits wide with parity disabled.
29K bytes of non-volatile configuration memory.
8192K bytes of processor board System flash (Read/Write)

Configuration register is 0x2102

Output from sh run:

Current configuration : 1190 bytes
!
version 12.1
no service single-slot-reload-enable
service timestamps debug uptime
service timestamps log uptime
no service password-encryption
!
hostname xxxxxxxxx
!
no logging buffered
logging rate-limit console 10 except errors
no logging console
no logging monitor
enable secret 5 xxxxxxxxxxxxxxx
!
ip subnet-zero
ip cef
!
!
no ip finger
!
!
!
!
interface FastEthernet0/0
ip address 192.168.1.45 255.255.255.128
speed 100
half-duplex
!
interface FastEthernet0/1
ip address 192.168.2.1 255.255.255.0
speed 100
half-duplex
!
ip classless
ip route 0.0.0.0 0.0.0.0 192.168.1.1
no ip http server
!
!
line con 0
password xxxxxxx
logging synchronous
login
transport input none
line aux 0
line vty 0 4
password xxxxxxx
logging synchronous
login
!
end

I've tried turning on CEF, turning off CEF, using no ip route-cache
same-interface, Full Duplex and Half Duplex modes and I cannot get the
Cisco router to perform better than the Windows 2000 Server.

The router is currently setup simply with static routing so there is
little CPU overhead.

Anyone care to comment on possible reasons for the W2K box to
outperfom the Cisco?? I thought the 3600 series would be at least on
par.
I spoke to an engineer from a company that does work for BT and told
him of the planned usage of the 3620 on our LAN and he never mentioned
that it would lack in performance.

Any ideas would be greatly appreciated.

Regards

Andrew Smith

 

:We recently purchased a Cisco 3620 to replace a Windows 2000 Server.
:The only module that is loaded is a Dual Fast Ethernet NM-2FE2W.
:This is being used as a LAN router (servers one side -- workstations
ther).

:Windows 2000 Server
: Transfer Rate: 1.263 MB/s

:Cisco 3620 Router - Full duplex 100Mbit
: Transfer Rate: 0.237 MB/s

With that kind of transfer rate, I'd say that you almost certainly
have a duplex conflict.

Do not trust autonegotiation on Windows 2000: I have personally
seen Windows 2K Server lie about it's settings. It was claiming
that it had negotiated 100 full duplex, but my network probe showed it
was really only 100 half duplex. Forcing both ends of the link
to full-duplex cleared up that particular problem.

 

I agree. It is almost certainly an auto-negotiation issue. This is just
about the most common LAN problem on Earth. I think that often the
performance is not as bad as what you are seeing and measuring, so the
problem can go a while without being detected.


Best regards,

Tom Larus, CCIE #10,014
Author of CCIE Warm-Up: Advice and Learning Labs
http://www.ipexpert.com/products_services/product.asp?sku=ip7777

 

For starters, stop using file transfers to measure network
performance. It adds too many extra bottlenecks like disk, pci-bus,
windows crappy file server performance, etc. Try running ttcp or
ttcpw which just transfers raw data and gives better numbers. There
is a java version of ttcp that runs on any platform with java.

-Chris

 

Further to this I did use Netlatency's SwitchMonitor and it showed one
interface as HALF-DUPLEX. I had actually forced the Fast Ethernet
interface to full duplex and a 'show int' command showed both
interfaces as FULL-DUPLEX.

As SwitchMonitor is picking up this status via SNMP I just assumed
that SwitchMonitor got it wrong, as the Cisco was showing both
interfaces as FULL-DUPLEX.

I will check to see whether the switch port that the cisco interface
is connected to could possibly be faulty.

Can anyone confirm though that the Cisco 3620 would outperform a
Windows 2000 Server ?? I would of thought the 3600 series would at
least by on par (3620 = 40kpps vs W2K supposed 50kpps).

Thanks for your advice.

Andrew

 

Besides from the probable duplex mismatch you seem to have:

IMO that's rather simple when you indeed compare a pretty old non
distributed router chassis like the 3620 (which is mostly a WAN router
design) to a somehwat new PC with somewhat decent CPU and I/O bus setup
(lets say at least multiple PCI buses and non-gameware NICs, i.e. no
RTL 8139s, but Intel EEPro or Tulips). It's just a matter of brute force.

Routing two times 100Base Ethernet? I don't have numbers handy, but I
don't have too good a feeling with that. Newer members of the 26xx
family have better CPUs than the 3620 and might get a grip here. Or
have a closer look on the 366x - this is a complete redesign, years
newer than the 36[24]0.

Did you mention you are planning to route *Fast* Ethernet with this box
and expect line speed throughput?

Boxes that can route even significantly more than just two 100Base
interfaces would be the Cat3550 or even more charming the Cat3750
series. You will get almost all features the 3620 has with respect
to routing if chosing the EI images. But you don't get WAN. Then
again, you get a decent switch as a bonus

 

Ok, after performing a number of different tests I settled on using
netcps to benchmark raw tcp transfer speed.

By forcing both the FastEthernet interface on the Cisco 3620 router
and the switch port on my switch to half-duplex I was able to get 4.6
Megabytes/sec transfer speeds.

Forcing full duplex dropped the throughput down to an average of 0.6
Megabytes/sec.

Funny enough if I change both the switch AND the router to duplex auto
then transfer speeds do reach 4.6 Megabytes/sec but the switch reports
the interface to be FULL DUPLEX while the router reports HALF DUPLEX.
Data transfers without errors however.

If I force a mismatch then the router goes crazy with transfer errors.

Weird. But I'm happy with 4.6 Megabytes/sec for now.
This is much faster than the average 2.0 Megabytes/sec we were getting
with a Windows 2000 Server configured as a LAN router.

Cheers,
Andrew.