[c-nsp] Using ATM PCR and MBS Values to Reduce Large Packet Latency

[Mark]

Hiya,

I have been puzzling over how I can use the ATM PCR and MBS values to reduce
the latency of large packets over small ATM VCs wich are running well under
SCR and was hoping someone could help.

I am sending traffic via a smartbits tester to test a 64K ATM VC:

- Flow of 32,000 bps made up of 1500 byte (12,000 bit) packets over the
64,000bps ATM VC

interface ATM6/0.200 point-to-point
 description 64k PVC
 ip address 2.2.2.2 255.255.255.252
 pvc 0/100
  protocol ip 2.2.2.1 broadcast
  vbr-nrt 71 71 1
  broadcast
  tx-ring-limit 3
  oam-pvc manage 5
  encapsulation aal5snap
  max-reserved-bandwidth 85

Key values are "vbr-nrt <PIR -Kbps> <CIR - Kbps> <MBS - cells>"

PCR: 71 Kbps
SCR: 71 Kbps
MBS:1

The SCR and PCR are 71 rather than 64 to allow for the 54/48 overhead of ATM
cell headers MBS is 1 as there is no burst configured.

The idea of not configuring a burst and reducing the tx-ring limit is that I
will be running layer 3 QOS over the PVC and I want that to take care of
controlling traffic rather than having the VC PCR burst (and resultant CLP=1
cells) do this.

When I send 1500 byte packets at 32Kbps the average latency per packet is
around 200ms.

Looking at Cisco's GCRA documentation, this makes sense. You have to acquire
tokens in order to be able to transmit over a VC.

It takes 12000 (packet size) / 64000 ( bps ATM SCR) = 0.1875 = 187.5ms to
transmit a 1500 byte packet over a 64K VC, even though the packets are
arriving at a rate only half that of the VC SCR

In comparison, sending the same 1500 byte packet flow at 32Kbps over a
64Kbps Frame Relay PVC the latency is around 10ms - the time it takes to
clock the packets onto the actual E1 circuit. Bring back Frame Relay, no
really :)

After all that blurb, the question:

How can I tune the PCR and MCR values to reduce large packet latency, but
without scuppering all hope of running L3 QOS effectively?

Telco has provided the following values for the VC configured on their
switch:

MBS: 250 cells
IBS: 63 cells
CDVT: 50,000 usec
CLP Threshold: 90%
No value for PCR given.

Eg if I configure the following VC values latency comes down to 37ms on
average, but this was through experimentation, I was hoping there would be a
way to calculate the optimum values.

pvc 0/100
  vbr-nrt 500 71 500