[LARTC] CPU speed vs. QoS

[Amit Kucheria <amitk@ittc.ku.edu>]

Hi,

We are testing some hybrid Diffserv-Intserv architectures and due to 
problems with some testing, we eventually tried to do a very simple test - 

Send traffic between the two machines (M1, M2) with an ingress policer on
the destination machine.

However there is a catch: 
1. M1 is a high speed machine (PIII, 1 Ghz, 1Gb RAM)
2. M2 is a low speed machine (PPro, 200Mhz, 128 Mb RAM)

We are sending about 7.8 Mbps of traffic with 1472 byte packets (we are 
using realtime patches to Linux for this)

This is our policer config string:
$TC filter add dev $DEVICE parent ffff: protocol ip prio 50 rsvp ipproto
udp session 129.237.125.147/9001 police rate 1043872bps burst 1515 drop
flowid :1

Scenario 1: M2 sending to M1
----------------------------
Many packets (22%) are marked as 'out of profile' and dropped by the 
policer on M1

Scenario 2: M1 sending to M2
----------------------------
M2 recieves about 98% packets, but those that are recieved are 
'in-profile'

These results are a little puzzling, since in Scenario 1, if anything, we
would have expected M2 to be able to process everything quickly.
And in Scenario 2, if anything, we would have expected out-of-profile 
packets since M2 would get swamped (being a slower machine)

Can anybody explain reasons for this behaviour? Anybody with benchmarks on 
CPU/Mem usage for different kinds of traffic?

Regards,
Amit

-- 
I'm an angel!!! Honest! 
The horns are just there to hold the halo up straight.
^-^-^-^-^-^-^-^-^-^-^-^-^-^-^-^-^-^-^-^-^-^-^-^-^-^-^
                  Amit Kucheria
          EECS Grad. Research Assistant
         University of Kansas @ Lawrence
   (R)+1-(785)-830 8521 ||| (O)+1-(785)-864 7774
____________________________________________________

_______________________________________________
LARTC mailing list / LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/