How big is 1 network packet?

[Tim D Fulford]

I have a network interface that is getting some 6000 packets per second. I'm pritty sure it is 100Mbit/s card.

My MTU is 1500, I believe this is in bits + a header. so say 2kbits this would give 11.7 Mbits/s all fine, and it is a max value.

BUT some of my collegues are saying it is much larger, closer to 1kByte this gives 46.9 Mbits/s. This is quite a high value and would suggest we may need to start thinking about reducing the segmet size to keep collisions down.

Tim

[Steven Gillard_2]

Tim,

The MTU is in octets, not bits, so your colleagues are correct. The MTU is the maximum payload of a link level frame, which is 1500 on most 100baseT networks. This limits you to 1460 octets of "real" data in TCP/IP packets because of the TCP and IP headers.

Of course, its unlikely that most of your packets are this large, so it would be wrong to simply multiply your inbound packet rate by the MTU to get a figure in octets/sec.

Regards,
Steve

[Bill Hassell]

Also note that if you are running your 100 Mbit card in full duplex (recommended) then there is no such thing as a collision. Collisions are possible only in half-duplex. Use lanscan to determine your card ID's then run lanadmin -x to determine the current speed and duplex. NOTE: 10.20 and earlier use CrdIn# from lanscan, 11.0 and later use NMID.

If your switch (I would not recommend simple hubs for 100Mbit networks) is set for auto-negotiation as well as your HP 9000, check to see that full-duplex has been negotiated. IF not, either a combination of chip vendors and/or cable length prevented accurate negotiation and you will get terrible (less than 5 Mbit/s) throughput with frame errors and collisions. To fix this, change BOTH your switch and the 9000 to 100, full duplex, manual.

[Tim D Fulford]

Thanks for the above. I suppose I really should have staed awake during those network lectures!

Bill - The interface is set up as half duplex. It is conected directly to an hp pro-curve hub. I'm told this is because the hub is auto... & has 100Mb/s (my card) on one side conecting a 10Mb/s on the other. The "other side" is a bit of an unknown as it my well be segmented off by a switch (I hope this is the case as 46Mb/s would flood it if it were part of the same segment).

Steve - Many thanks

anyone - Is there a way of finding out the bandwidth in MeasureWare

Thanks for your help

Tim

[Steven Gillard_2 · Accepted solution]

Doesn't look like mwa has such a metric on HPUX, the only ones documented are:

# grep NET /var/opt/perf/reptfile
* GBL_NUM_NETWORK
GBL_NET_IN_PACKET_RATE
GBL_NET_OUT_PACKET_RATE
* GBL_NET_PACKET_RATE
GBL_NET_ERROR_1_MIN_RATE
* GBL_NET_COLLISION_1_MIN_RATE
GBL_NET_IN_ERROR_PCT
GBL_NET_OUT_ERROR_PCT
GBL_NET_OUTQUEUE
GBL_NET_COLLISION_PCT
GBL_NETWORK_SUBSYSTEM_QUEUE
APP_NETWORK_SUBSYSTEM_WAIT_PCT
DATA TYPE NETIF
BYNETIF_NAME
BYNETIF_IN_PACKET_RATE
BYNETIF_OUT_PACKET_RATE
BYNETIF_ERROR_RATE
BYNETIF_COLLISION_RATE
GBL_NUM_NETWORK

Which is pretty much everything you get in glance.

Attached is a rather crude perl script that will give you an approximate kbit/sec reading by using the lanadmin stats. Run it as follows:

# ./netperf.pl

I'm sure there's some 3rd party solutions out there that will give you much more accurate information. You're likely to get better information off the switch as well.

Regards,
Steve

[Mark van Hassel]

Tim,

Bill stated:

"NOTE: 10.20 and earlier use CrdIn# from lanscan, 11.0 and later use NMID. "

It is just the other way around: 10.20 and earlier require the NMID fromm lanscan,HPUX 11.0 and later require the PPA or CrdIn number.

Be carefull with this since entering a non -existing number will give you no error message.

Good luck,

Mark.

[Tim D Fulford]

Many many thanks for the perl script. As it happens 1 packet is about 1kbit!!!! So I was right for the wrong reasons!

However, a furthur 10 points can be awarded if you can get your perl script to calculate

kbits Pkt_Rate Pkt_Size

Just a carrot, if you can't the script is super

many thanks to Steve & Bill, you have given me more to think about!

Cheers

Tim

[Steven Gillard_2]

OK, but only because I'm bored :)

The new script is attached, it includes an approximation of packets/sec and bytes/packet (which is an average). Note that these are only the inbound stats the script is looking at, if you want outbound stats you can change the get_data routine to look at the "Outbound" lanadmin values instead of the "Inbound" values.

Cheers,
Steve