Experiment
on Oct 6
Conclusion: 1. linux kernel 2.4.10 is much faster than kernel
2.2.16-22 in ip forwarding
2. 3com
card is a little faster than DEC Tulip
To get a pure
text version of experiment data page, see here
Dublin and prague are
running kernel 2.4.10
(NOTE: This time, the tulip card driver is built into kernel directly )
madrid is running kernel
2.2.16
Strange things:
When I boot kernel 2.2.16-22, the Ethernet interfaces information:
3Com:eth0 tulip1:eth1 tulip2:eth2
When I boot kernel 2.4.10, the Ethernet interfaces information:
tulip1:eth0 tulip2: eth1 3Com:eth2
What's going on here?????? How can I bind 3Com to eth0
permanently????
|
Routing Experiment 2 on
Kernel 2.4.10 |
| NICs involved in routing |
3Com + Tulip |
Tulip + Tulip |
|
| |
1st |
2nd |
3rd |
4th |
5th |
6th |
7th |
AVG |
| Trans Speed (KB/sec) |
7506.90 |
7822.8 |
7644.09 |
7818.16 |
7866.14 |
7804.45 |
7752.20 |
7744.963 |
| Recv Speed (KB/sec) |
7469.57 |
7795.41 |
7630.09 |
7806.88 |
7855.23 |
7793.61 |
7760.29 |
7730.154 |
| Data Sent (Byte) |
16777216 |
16777216 |
16777216 |
16777216 |
16777216 |
16777216 |
16777216 |
|
|
Routing
Experiment 2 on Kernel 2.2.16-22 |
| |
1st |
2nd |
3rd |
AVG |
| Trans Speed (KB/sec) |
252.78 |
251.84 |
250.85
|
251.8233 |
| Recv Speed (KB/sec) |
251.92 |
251.07 |
250.02 |
251.0033 |
| Data Sent (Byte) |
16777216 |
16777216 |
16777216 |
|
From the above data and the
data from last week, we can conclude
that:
linux kernel 2.4.10 is much faster than kernel
2.2.16-22 in ip forwarding
For the experiment data detail, see the following:
=================================================================
Experiment 1:
[lin@prague ttcp]$ ttcp -r -s
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
ttcp-r: accept from 192.168.3.1
ttcp-r: 16777216 bytes in 1.44 real seconds = 11370.92
KB/sec +++
ttcp-r: 12285 I/O calls, msec/call = 0.12, calls/sec = 8526.11
ttcp-r: 0.0user 0.2sys 0:01real 18% 0i+0d 0maxrss 0+1pf 0+0csw
[lin@madrid lin]$ ttcp/ttcp -t -s 192.168.3.2 <
nttcp-1.47.tar.gz
ttcp-t: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp -> 192.168.3.2
ttcp-t: socket
ttcp-t: connect
ttcp-t: 16777216 bytes in 1.43 real seconds = 11460.60
KB/sec +++
ttcp-t: 2048 I/O calls, msec/call = 0.71, calls/sec = 1432.58
ttcp-t: 0.0user 0.0sys 0:01real 4% 0i+0d 0maxrss 0+2pf 0+0csw
=================================================================
Experiment 2:
[lin@prague ttcp]$ ttcp -r -s
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
ttcp-r: accept from 192.168.1.2
ttcp-r: 16777216 bytes in 2.19 real seconds =
7469.57 KB/sec +++
ttcp-r: 12285 I/O calls, msec/call = 0.18, calls/sec = 5600.81
ttcp-r: 0.0user 0.1sys 0:02real 9% 0i+0d 0maxrss 0+1pf 0+0csw
[lin@madrid lin]$ ttcp/ttcp -t -s 192.168.2.2 <
nttcp-1.47.tar.gz
ttcp-t: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp -> 192.168.2.2
ttcp-t: socket
ttcp-t: connect
ttcp-t: 16777216 bytes in 2.18 real seconds =
7506.90 KB/sec +++
ttcp-t: 2048 I/O calls, msec/call = 1.09, calls/sec = 938.36
ttcp-t: 0.0user 0.1sys 0:02real 4% 0i+0d 0maxrss 0+2pf 0+0csw
=================================================================
Experiment 3. USING SWITCH
[lin@prague ttcp]$ ttcp -r -s
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
ttcp-r: accept from 132.177.8.27
ttcp-r: 16777216 bytes in 1.96 real seconds =
8374.19 KB/sec +++
ttcp-r: 12246 I/O calls, msec/call = 0.16, calls/sec = 6259.18
ttcp-r: 0.0user 0.2sys 0:01real 11% 0i+0d 0maxrss 0+2pf 0+0csw
[lin@madrid lin]$ ttcp/ttcp -t -s 132.177.8.29 <
nttcp-1.47.tar.gz
ttcp-t: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp -> 132.177.8.29
ttcp-t: socket
ttcp-t: connect
ttcp-t: 16777216 bytes in 1.95 real seconds =
8407.53 KB/sec +++
ttcp-t: 2048 I/O calls, msec/call = 0.97, calls/sec = 1050.94
ttcp-t: 0.0user 0.2sys 0:01real 13% 0i+0d 0maxrss 0+2pf 0+0csw
From the above 3 experiments:
We found the networking speed improves significantly,
compared to last week's experiment which is
running kernel 2.2.16-22 on Dublin
The possible reason may be:
1. Dublin is using new Kernel 2.4.10 which may improve its routing
algorithm or something
To test this assumption, reboot prague.cs.unh.edu to kernel 2.2.16 , which
means the end computers are the same as those in
last week's experiment.
Do the same experiment 2 for 3 times
============================================================
[lin@prague ttcp]$ ttcp -r -s
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
ttcp-r: accept from 192.168.1.2
ttcp-r: 16777216 bytes in 2.10 real seconds =
7795.41 KB/sec +++
ttcp-r: 11672 I/O calls, msec/call = 0.18, calls/sec = 5553.47
ttcp-r: 0.0user 0.0sys 0:02real 1% 0i+0d 0maxrss 0+2pf 0+0csw
[lin@madrid lin]$ ttcp/ttcp -t -s 192.168.2.2 <
nttcp-1.47.tar.gz
ttcp-t: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp -> 192.168.2.2
ttcp-t: socket
ttcp-t: connect
ttcp-t: 16777216 bytes in 2.09 real seconds =
7822.83 KB/sec +++
ttcp-t: 2048 I/O calls, msec/call = 1.05, calls/sec = 977.85
ttcp-t: 0.0user 0.0sys 0:02real 2% 0i+0d 0maxrss 0+2pf 0+0csw
============================================================
============================================================
[lin@prague ttcp]$ ttcp -r -s
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
ttcp-r: accept from 192.168.1.2
ttcp-r: 16777216 bytes in 2.15 real seconds =
7630.09 KB/sec +++
ttcp-r: 11689 I/O calls, msec/call = 0.19, calls/sec = 5443.61
ttcp-r: 0.0user 0.0sys 0:02real 1% 0i+0d 0maxrss 0+1pf 0+0csw
[lin@madrid lin]$ ttcp/ttcp -t -s 192.168.2.2 <
nttcp-1.47.tar.gz
ttcp-t: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp -> 192.168.2.2
ttcp-t: socket
ttcp-t: connect
ttcp-t: 16777216 bytes in 2.14 real seconds =
7644.09 KB/sec +++
ttcp-t: 2048 I/O calls, msec/call = 1.07, calls/sec = 955.51
ttcp-t: 0.0user 0.0sys 0:02real 2% 0i+0d 0maxrss 0+2pf 0+0csw
============================================================
============================================================
[lin@prague ttcp]$ ttcp -r -s
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
ttcp-r: accept from 192.168.1.2
ttcp-r: 16777216 bytes in 2.10 real seconds =
7806.88 KB/sec +++
ttcp-r: 11681 I/O calls, msec/call = 0.18, calls/sec = 5565.92
ttcp-r: 0.0user 0.0sys 0:02real 0% 0i+0d 0maxrss 0+2pf 0+0csw
[lin@madrid lin]$ ttcp/ttcp -t -s 192.168.2.2 <
nttcp-1.47.tar.gz
ttcp-t: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp -> 192.168.2.2
ttcp-t: socket
ttcp-t: connect
ttcp-t: 16777216 bytes in 2.10 real seconds =
7818.16 KB/sec +++
ttcp-t: 2048 I/O calls, msec/call = 1.05, calls/sec = 977.27
ttcp-t: 0.0user 0.0sys 0:02real 2% 0i+0d 0maxrss 0+2pf 0+0csw
============================================================
HERE we see that :
if the 3COM card is in the same speed as tulip card,
the new kernel 2.4.10 DOES improve the routing
speed significantly
( 7818.16
KB/sec +++ VS 230KB/sec ++++
===> 34:1 )
We still need to remove the possibility that this improvement is brought
by the speed difference of 3COM card and tulip card.
In the following 3 experiments. We change NIC setting on dublin. USE tulip
card as NIC involed in static routing between 192.168.1.0 and 192.168.2.0
Do experiment 2 for 3 times
==========================================================
[lin@prague ttcp]$ ttcp -r -s
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
ttcp-r: accept from 192.168.1.2
ttcp-r: 16777216 bytes in 2.09 real
seconds = 7855.23 KB/sec +++
ttcp-r: 11695 I/O calls, msec/call = 0.18, calls/sec = 5607.11
ttcp-r: 0.0user 0.0sys 0:02real 3% 0i+0d 0maxrss 0+1pf 0+0csw
[lin@madrid lin]$ ttcp/ttcp -t -s 192.168.2.2 <
nttcp-1.47.tar.gz
ttcp-t: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp -> 192.168.2.2
ttcp-t: socket
ttcp-t: connect
ttcp-t: 16777216 bytes in 2.08 real
seconds = 7866.14 KB/sec +++
ttcp-t: 2048 I/O calls, msec/call = 1.04, calls/sec = 983.27
ttcp-t: 0.0user 0.0sys 0:02real 3% 0i+0d 0maxrss 0+2pf 0+0csw
==========================================================
==========================================================
[lin@prague ttcp]$ ttcp -r -s
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
ttcp-r: accept from 192.168.1.2
ttcp-r: 16777216 bytes in 2.10 real
seconds = 7793.61 KB/sec +++
ttcp-r: 11729 I/O calls, msec/call = 0.18, calls/sec = 5579.30
ttcp-r: 0.0user 0.0sys 0:02real 2% 0i+0d 0maxrss 0+1pf 0+0csw
[lin@madrid lin]$ ttcp/ttcp -t -s 192.168.2.2 <
nttcp-1.47.tar.gz
ttcp-t: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp -> 192.168.2.2
ttcp-t: socket
ttcp-t: connect
ttcp-t: 16777216 bytes in 2.10 real
seconds = 7804.45 KB/sec +++
ttcp-t: 2048 I/O calls, msec/call = 1.05, calls/sec = 975.56
ttcp-t: 0.0user 0.1sys 0:02real 4% 0i+0d 0maxrss 0+2pf 0+0csw
==========================================================
==========================================================
[lin@prague ttcp]$ ttcp -r -s
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
ttcp-r: accept from 192.168.1.2
ttcp-r: 16777216 bytes in 2.11 real
seconds = 7752.20 KB/sec +++
ttcp-r: 11703 I/O calls, msec/call = 0.18, calls/sec = 5537.35
ttcp-r: 0.0user 0.0sys 0:02real 4% 0i+0d 0maxrss 0+1pf 0+0csw
[lin@madrid lin]$ ttcp/ttcp -t -s 192.168.2.2 <
nttcp-1.47.tar.gz
ttcp-t: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp -> 192.168.2.2
ttcp-t: socket
ttcp-t: connect
ttcp-t: 16777216 bytes in 2.11 real
seconds = 7760.29 KB/sec +++
ttcp-t: 2048 I/O calls, msec/call = 1.06, calls/sec = 970.04
ttcp-t: 0.0user 0.0sys 0:02real 4% 0i+0d 0maxrss 0+2pf 0+0csw
==========================================================
In the above 3 experiment, on Dublin the NIC cards involved
in routing are 2 tulip cards.
Hence, we can draw a conclusion that:
Kernel 2.4.10 is significantly faster than kernel 2.2.16-22 on
static routing
Reboot Dublin to 2.2.16-22 to prove that.
Do experiments 2 for 3 times
=================================================
[lin@prague ttcp]$ ttcp -r -s
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
ttcp-r: accept from 192.168.1.2
ttcp-r: 16777216 bytes in 65.04 real seconds =
251.92 KB/sec +++
ttcp-r: 5154 I/O calls, msec/call = 12.92, calls/sec = 79.25
ttcp-r: 0.0user 0.0sys 1:05real 0% 0i+0d 0maxrss 0+2pf 0+0csw
[lin@madrid lin]$ ttcp/ttcp -t -s 192.168.2.2 <
nttcp-1.47.tar.gz
ttcp-t: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp -> 192.168.2.2
ttcp-t: socket
ttcp-t: connect
ttcp-t: 16777216 bytes in 64.82 real seconds =
252.78 KB/sec +++
ttcp-t: 2048 I/O calls, msec/call = 32.41, calls/sec = 31.60
ttcp-t: 0.0user 0.0sys 1:04real 0% 0i+0d 0maxrss 0+2pf 0+0csw
=================================================
=================================================
[lin@prague ttcp]$ ttcp -r -s
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
ttcp-r: accept from 192.168.1.2
ttcp-r: 16777216 bytes in 65.26 real seconds =
251.07 KB/sec +++
ttcp-r: 5164 I/O calls, msec/call = 12.94, calls/sec = 79.13
ttcp-r: 0.0user 0.0sys 1:05real 0% 0i+0d 0maxrss 0+2pf 0+0csw
[lin@madrid lin]$ ttcp/ttcp -t -s 192.168.2.2 <
nttcp-1.47.tar.gz
ttcp-t: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp -> 192.168.2.2
ttcp-t: socket
ttcp-t: connect
ttcp-t: 16777216 bytes in 65.06 real seconds =
251.84 KB/sec +++
ttcp-t: 2048 I/O calls, msec/call = 32.53, calls/sec = 31.48
ttcp-t: 0.0user 0.0sys 1:05real 0% 0i+0d 0maxrss 0+2pf 0+0csw
=================================================
=================================================
[lin@prague ttcp]$ ttcp -r -s
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
ttcp-r: accept from 192.168.1.2
ttcp-r: 16777216 bytes in 65.53 real seconds =
250.02 KB/sec +++
ttcp-r: 5157 I/O calls, msec/call = 13.01, calls/sec = 78.70
ttcp-r: 0.0user 0.0sys 1:05real 0% 0i+0d 0maxrss 0+2pf 0+0csw
[lin@madrid lin]$ ttcp/ttcp -t -s 192.168.2.2 <
nttcp-1.47.tar.gz
ttcp-t: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp -> 192.168.2.2
ttcp-t: socket
ttcp-t: connect
ttcp-t: 16777216 bytes in 65.31 real seconds =
250.85 KB/sec +++
ttcp-t: 2048 I/O calls, msec/call = 32.66, calls/sec = 31.36
ttcp-t: 0.0user 0.0sys 1:05real 0% 0i+0d 0maxrss 0+2pf 0+0csw
=================================================
So kernel 2.4.10 is much faster than kernel 2.2.16
3. From experiment 3
madrid:eth0( 3 COM ) <===> prague:eth0( tulip ),
we see that tulip seems slower then COM.( compared to
last week experiment
which is 3COM <====> 3COM. 8407.53 KB/sec VS 11461.40 KB/sec +++)
To test this assumption, reboot Prague to kernel 2.2.16 which will use 3COM
as eth0. Re-do experiment 3 to see the speed.
Do the same experiment 3 for 3 times
===================================================
[lin@prague ttcp]$ ttcp -r -s
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
ttcp-r: accept from 132.177.8.27
ttcp-r: 16777216 bytes in 1.44 real seconds =
11409.07 KB/sec +++
ttcp-r: 12275 I/O calls, msec/call = 0.12, calls/sec = 8547.75
ttcp-r: 0.0user 0.2sys 0:01real 18% 0i+0d 0maxrss 0+1pf 0+0csw
[lin@madrid lin]$ ttcp/ttcp -t -s 132.177.8.29 <
nttcp-1.47.tar.gz
ttcp-t: buflen=8192,nbuf=2048, align=16384/0, port=5001 tcp -> 132.177.8.29
ttcp-t: socket
ttcp-t: connect
ttcp-t: 16777216 bytes in 1.43 real seconds =
11448.47 KB/sec +++
ttcp-t: 2048 I/O calls, msec/call = 0.72, calls/sec = 1431.06
ttcp-t: 0.0user 0.3sys 0:01real 26% 0i+0d 0maxrss 0+2pf 0+0csw
===================================================
===================================================
[lin@prague ttcp]$ ttcp -r -s
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
ttcp-r: accept from 132.177.8.27
ttcp-r: 16777216 bytes in 1.44 real seconds =
11395.89 KB/sec +++
ttcp-r: 12245 I/O calls, msec/call = 0.12, calls/sec = 8517.01
ttcp-r: 0.0user 0.1sys 0:01real 11% 0i+0d 0maxrss 0+2pf 0+0csw
[lin@madrid lin]$ ttcp/ttcp -t -s 132.177.8.29 <
nttcp-1.47.tar.gz
ttcp-t: buflen=8192,nbuf=2048, align=16384/0, port=5001 tcp -> 132.177.8.29
ttcp-t: socket
ttcp-t: connect
ttcp-t: 16777216 bytes in 1.43 real seconds =
11454.46 KB/sec +++
ttcp-t: 2048 I/O calls, msec/call = 0.72, calls/sec = 1431.81
ttcp-t: 0.0user 0.3sys 0:01real 21% 0i+0d 0maxrss 0+2pf 0+0csw
===================================================
===================================================
[lin@prague ttcp]$ ttcp -r -s
ttcp-r: buflen=8192, nbuf=2048, align=16384/0, port=5001 tcp
ttcp-r: socket
ttcp-r: accept from 132.177.8.27
ttcp-r: 16777216 bytes in 1.44 real seconds =
11413.35 KB/sec +++
ttcp-r: 12284 I/O calls, msec/call = 0.12, calls/sec = 8557.23
ttcp-r: 0.0user 0.1sys 0:01real 13% 0i+0d 0maxrss 0+1pf 0+0csw
[lin@madrid lin]$ ttcp/ttcp -t -s 132.177.8.29 <
nttcp-1.47.tar.gz
ttcp-t: buflen=8192,nbuf=2048, align=16384/0, port=5001 tcp -> 132.177.8.29
ttcp-t: socket
ttcp-t: connect
ttcp-t: 16777216 bytes in 1.43 real seconds =
11453.28 KB/sec +++
ttcp-t: 2048 I/O calls, msec/call = 0.72, calls/sec = 1431.66
ttcp-t: 0.0user 0.5sys 0:01real 39% 0i+0d 0maxrss 0+2pf 0+0csw
===================================================
HERE we do see that 3COM <===> 3COM is a little faster than 3COM <==>
tulip
To draw the conclusion, we need more experiments