IRQ, network drivers, and the like...

Peter Van Epp vanepp at sfu.ca
Tue Sep 25 14:59:31 EDT 2001


	1 interrupt per packet would be my impression too. Are you sure of the 
input packet rate (i.e. is the packet count the same from an rmon capable 
network device connected to the same network)? Things to look at would be DNS 
lookups being done per packet (such as if ra is reading the data stream as it 
rolls by) from the same network, and argus running out the same interface as a
network server (although the packet rate from that should be much lower 
that the input stream rate). Something may be generating extra traffic you
haven't accounted for. A tcpdump from a second machine would be interesting
to see what and how many packets it thinks are being delivered (or an rmon
capable device as I said). We always use 3c509s, I don't know if Martin has
experimented with 3c59x cards. 3 3c509s in a dual processor machine (with
I think Mr Becker's link aggregating drivers) can get 294 megs per second
between machines in a beowolf cluster. That eats one entire CPU for the 
network between interrupts and stack processing leaving the other CPU for
doing useful work so you just need to accept that the net is likely to need
to eat an entire CPU and not much is going to help that. The 3c905s are the
fastest cards we have found (the Etherexpress is a small fraction slower).

Peter Van Epp / Operations and Technical Support 
Simon Fraser University, Burnaby, B.C. Canada


> 
> Yup, 30k interrupts/second.
> 
>   Well, I had a brief discussion with Donald Becker, who had written most of 
> the linux network drivers, and he thought this was weird, and indicated a 
> problem.  He thought the card(s) I was using should generate slightly less 
> then the number of TX/RX packets/second.
> 
>   Problem is, after sending him some info, he seems to have dropped away for a 
> while.  Likely on vacation or something... mmmmm.... vacation :)
> 
>   What are other people's experiences with the 3com 3c59x class of cards?  
> Does 30K interrupts sound reasonable for 10K packets?
> 
>   Is this high?  IE: is this affecting system performance for other devices?  
> Should I invest in a better card(s) with interrupt coallecing (sp?)?
> 
> Thanks
> 
> Chris
> 
> 
> >===== Original Message From <carter at qosient.com> =====
> >Hey Chris,
> >   Do you mean interrupts per second?  I can understand how the
> >card could generate (3 x (the number of packet)) interrupts.
> >Start of bus transfer, end of bus transfer, and specific device
> >handler interrupt, but this is just a guess, I don't really
> >know.  Possibly Peter will have a real educated guess as to
> >what is going on, and knowing him, possibly a way around it.
> >
> >Carter
> >
> >Carter Bullard
> >QoSient, LLC
> >300 E. 56th Street, Suite 18K
> >New York, New York  10022
> >
> >carter at qosient.com
> >Phone +1 212 588-9133
> >Fax   +1 212 588-9134
> >http://qosient.com
> >
> >> -----Original Message-----
> >> From: Chris Newton [mailto:newton at unb.ca]
> >> Sent: Tuesday, September 25, 2001 1:57 PM
> >> To: carter at qosient.com
> >> Subject: IRQ, network drivers, and the like...
> >>
> >>
> >> Hi Carter,  how have you been?
> >>
> >>   I'm struggling with a weird problem and I thought maybe you
> >> might have seen
> >> this before, or be able to point me in the right direction.
> >>
> >>   One of the networks I'm watching (my big network), has (for
> >> example), about
> >> 5000 incoming packets and 5000 outgoing right now... works
> >> out to about 2MB in
> >> both directions.  That is fed into an argus box, via 1
> >> cable... so, 10K
> >> packets per second on the cable.
> >>
> >>   Whats weird.. is that right now, my network card is
> >> generating about 26 to
> >> 30,000 packets/second.  I feel this is bogging the machine...
> >> and I dont know
> >> why it might be doing that.  Argus is running, reading from
> >> that interface.
> >>
> >>   The card is a 3Com, 3c980 (or something)... but, Im not
> >> sure that matters...
> >> because I swapped it for a intel etherner express pro, and I
> >> had the same
> >> problem... (even higher actually).  There is also an ra
> >> client reading from
> >> the localargus stream.
> >>
> >>   Weird.. but, do you have any ideas?
> >>
> >>
> 
> 



More information about the argus mailing list