RE: [RPRWG] Some questions about the pseudo-code in Gandalf
Anoop,
I am trying to give answers that can make you happy, though I recognize
that anything short of Alladin as a whole wouldn't make you complete happy.
Happy New Year !
Donghui
At 11:25 AM 1/2/2002 -0800, Anoop Ghanwani wrote:
>Donghui,
>
>Many thanks for response. I'm going to go ahead and
>delete the answers that I'm happy with, but would like
>to continue the discussion on some of the things that
>are still unclear in my mind.
>
>-Anoop
>
>
> > -----Original Message-----
> > From: Donghui Xie [mailto:dxie@xxxxxxxxx]
> > Sent: Wednesday, January 02, 2002 11:01 AM
> > To: Anoop Ghanwani
> > Cc: stds-802-17@xxxxxxxx
> > Subject: RE: [RPRWG] Some questions about the pseudo-code in Gandalf
> >
> > >- lp_allow is computed but never used. Is something missing?
> >
> > Please see the pseudo code in the section of "THESE ARE
> > UPDATED EVERY CLOCL
> > CYCLE" on page 46.
>
>In an actual implementation, would one really have to do these
>updates _every single_ clock tick? Or is there some way to make
>it less frequent? If so, what frequency of update is recommended?
The pseudo code simply states that token should be added to the dynamic
shaper bucket continuously. There are many implementations that can
effectively achieve the same result without adversely impeding intended
traffic shaping and transmission. I would leave system designers to
recommend a particular implementation as fit to their system design.
> > > > - What's the rationale for:
> > > > if (recd_rate != NULL_RCVD_INFO) &&
> > > > (lp_forward_rate > (allow_rate_congestion/WEIGHT))
> > > > in order to determine whether to send a rev_rate?
> >
> > This is a part of the fairness algorithm to ensure fair rate
> > transmission
> > on the ring. If some downstream node is congested and
> > upstream nodes are
> > sending more traffic than the fair rate of the current node
> > (in a weighted
> > fashion), current node will pass its received fair rate
> > further upstream.
> > This fairness logic applies when current node is not congested.
>
>The reason I brought this up is because only the local node's
>weight is being used to determine whether or not to send the
>message upstream. A better design, IMHO, would account for
>the weights of the other nodes as well.
I see much more advantage in using local weight only, which simplifies
weight calculation requirements while providing elegant weighted fairness
at the same time eliminates the need for global weights information and its
associated much more complex calculations. That said, in recognizing
Alladin's needs for global weights, next Gandalf version may optionally
allow people use global weights in their preferred hard way.
> >
> > > >
> > > > - The pseudo-code does not appear to account for virtual
> > > > output queueing support within the MAC client.
> >
> > The pseudo-code presents the basic fairness algorithm with single
> > choke-point information. A VoQ capable MAC client would run
> > multiple copies
> > of the basic fairness algorithm, and keep track of all the
> > choke-points
> > information with each running copy corresponding to one
> > virtual destination
> > queue. For details, please see section 6.7 for descriptions on the
> > relationship of the basic fairness algorithm and VoQ capable
> > MAC client.
>
>Section 6.7 does indeed explain the multi-choke point case.
>In the absence of multi-choke point capability, the algorithm
>can be shown to be shown to have severe limitations.
>Therefore, I think it makes sense to have the pseudo-code
>cover the multi-choke point case as well, even if it
>just involves adding a loop. Leaving it as an exercise for
>the interested reader can hide complexity that it
>introduces. I would like to see how it affects data
>traffic making its way from the MAC client to the MAC.
Claim of "severe limitations" and the "absence of multi-choke point
capability" are not substantiated. So no points are given on that. "hide
complexity" seems to contradict "Section 6.7 does indeed explain the
multi-choke point case."
802.17 MAC should be able to provide all the traffic information to MAC
client, which may choose to utilize the info and make intelligent packet
transmission so as to optimize ring utilization. Gandalf bandwidth
management enables all the means for a highly sophisticated MAC client to
transmit packets onto the ring while optimizing ring bandwidth
utilization. If you like to see how MAC client implementations affect data
traffic "making its way from the MAC client to the MAC", you are welcome to
exercise the various MAC client implementations.
> > > > > - First off, it seems that the pseudo-code does not handle
> > > > > the multi-choke point feature discussed in the proposal.
> > > > > It would be nice to see this.
> >
> > ditto.
>
>I guess you agree with my statement above :-)
Certinaly not.
> >
> > > > >
> > > > > - When a fairness packet is received, there's some
> > > > > code that says:
> > > > > if (fairness_pkt.SA == my_SA) &&
> > > > > ((node_state == wrapped)) {
> > > > > ..
> > > > > } else {
> > > > > ..
> > > > > }
> >
> > One check condition is missing here. We'll update it Gandalf in next
> > version. The complete check should be
> > if ( (fairness_pkt.SA == my_SA) &&
> > ( (fairness_pkt.ring_id == my_ring_id) || ( node_state ==
> > wrapped) ) ) {
> > rcvd_rate = NULL_RCVD_INFO;
> > } else {
> > rcvd_rate = fairness_pkt.rate;
> > }
> >
> > The first check states that if a node receives a fair rate
> > originated from
> > itself in same ringlet, which means its rate is the smallest
> > around the
> > ring, it will treat the received rate as a NULL info so that
> > its allowed
> > fair rate will be able to ramp up.
> >
> > In case of the node is in wrap, if it receives on its outer
> > ring a fair
> > rate info transmitted from its inner ring, it understands
> > that the ring is
> > wrapped, both inner ringlet and outer ringlet become one ring
> > essentially,
> > so its rate is the smallest around the ring, it will treat
> > this case the
> > same way as in previous case.
> >
> > > > > Should we also be checking the RING ID in the fairness
> > > > > packet? Or is it proposing that we don't do any
> > > > > bandwidth management when the ring wraps?
> >
> > Ditto. Gandalf checks RING ID, and its fairness algorithm guarantees
> > bandwidth fairness management in all situations.
>
>OK. That's what it seemed like to me when I read it. I wanted
>to make sure that you did intend to check the ring ID as well.
>
> > > > >
> > > > > - Can one of the authors comment on the tolerance
> > > > > of the scheme to message loss? The fairness message is
> > > > > being sent hop-by-hop, and it seems like if it happens
> > > > > to get lost, those that don't see it will continue
> > > > > to increase their rates. If that is indeed the case,
> > > > > I think something needs to be done about robustness.
> >
> > Fair rate messages get transmitted in fixed interval
> > periodically. The
> > interval is about 100us. In addition, Gandalf incorporates fair rate
> > ramp-down and ramp-up in a low-pass filtered fashion at the
> > intervals.
> > Overall, these schemes integrate and establish a very robust Gandalf
> > fairness algorithm, which is immune to occasional fair rate
> > message losses.
> > However, a persistent fair rate message loss over several consecutive
> > intervals will trigger Gandalf protection procedure.
>
>I thought that the fairness messages are sent no more often than
>each DECAY_INTERVAL. I haven't done the math, but maybe that does
>work out to 100 usec. The low pass filtered ramp-up is good.
>The low pass filtered ramp-down may not necessarily be the
>best in times of message loss because it doesn't give up
>bandwidth as quickly as it should (or could if the congestion
>messages are getting lost). I think simulations should consider
>the effect of fairness message loss on performance.
>(I'm pushing for quantifying behavior in worst-case scenarios
>as alluded to by other folks on the simulation thread.)
If you like, Gandalf doesn't stop you from sending more often fairness
messages. Actually, Gandalf doesn't do low pass in ramp-down, so Gandalf
may well be the best.
As Gandalf performance is not sensitive to occasional congestion message
loss, I don't see the necessity to spend much effort to study its effects.
Nonetheless, you are very much welcome to study the situations. I would
certainly be interested to see your results if you may kindly share them.
> >
> > > > >
> > > > > - What is the significance of the parameter AGECOEFF
> > > > > and how does one determine what is a good value to
> > > > > use for it?
> >
> > This parameter is part of Gandalf fair rate calculation,
> > which is used for
> > all transmission rates and not supposed to change at all.
>
>It's nice to know that I don't have to change it :-), but
>it doesn't help me understand why it is needed. Basically,
>I'd like to know why one should be so sure that 4 is best
>value.
Through heuristics and simulations, we made sure a value of 4 is the best.
Well, you are definitely welcome to try out simulations to be certain on
different values for it and share your discovery with .17 WG.