Re: [aqm] PIE vs. RED

["Rong Pan (ropan)" <[e-mail]>]

Please see inline.

Rong


On 8/14/15, 7:44 AM, "Francini, Andrea (Andrea)"
<[e-mail]> wrote:

>Hi Rong, 
>
>Thank you for pointing to Figure 8 in your HPSR 2013 paper.
>
>At first glance it exemplifies well the stability vs. variability of the
>equilibrium point established by the two schemes as the traffic intensity
>(number of concurrent flows) changes.
>
>However, there are two things I cannot explain after a closer look:
>
>1. At the very beginning the PIE curve shows a delay peak at about 240ms.
>With 2MB buffer size and 100Mbps rate, the maximum delay I would expect
>is 160ms (as indeed shown by the RED curve). Do you know where the extra
>80ms come from in the PIE case? It really looks like it is a single delay
>sample, so maybe it is just an issue with the initialization of the
>variables used in the delay measurements?

Yes, our queueing delay is calculated so there is an uptick in the
artificial delay at the beginning of the simulation when the draining rate
is just start to be calculated.



>
>2. If I understand correctly the RED configuration used for the plot,
>min_th is set at 32ms (20% of 2MB) and max_th is set at 128ms (80%). In
>the plot the RED delay rightly oscillates within this two bounds.
>However, if I compute the theoretical packet drop rate p using the
>formula  p = (1.22 * MSS / BW * RTT)^2 in the case with 50 flows
>(corresponding to the 100s-150s interval in the experiment) I get a drop
>rate value around 0.5%. Since you used a max_p of 0.1 (or 10%), your
>delay plot places the drop probability well above 5%. With the
>theoretical value of 0.5% from the formula the average delay would be
>only 1/20 of the way between min_th and max_th, or about 37ms, much lower
>than the almost 100ms of the plot. I ran an ns2 simulation with the
>parameters of the paper, getting steady-state average and maximum delays
>of 36ms and 47ms and a drop rate of 0.29% (with PIE I get 20.1ms, 39.4ms,
>and 0.38%). The only way I can explain the larger RED delay of your plot
>is with a lower value of max_p (e.g., with max_p = 1% instead of 10% my
>simulation yields 62.3ms and 71.9ms for the average and maximum delay,
>with 0.2% average drop rate; the delay is still not at the level of
>Figure 8, but getting closer). Note that with thresholds tuned around the
>20ms delay target of PIE (min_th = 16ms, max_th = 128ms) and max_p = 5%,
>the simulation yields Avg/Max delay of 17.8ms/29.5ms, pretty much the
>same as PIE. Overall, can you explain the discrepancy between expected
>and plotted delay for RED in Figure 8?

This is an omission in the paper, which we should specify. This simulation
is done with 10Mbps link with 200KB of buffer. We did size our buffers
according to the BDP. As Jonathan rightly pointed out, "the queue size is
tuned for the maximum capability of the device and a pessimistic value for
RTT². Figure 7 is intended to show that in reality link speed may not get
to the max capacity, and hence well-intended BDP buffer sizing may cause
extreme delays, the time between 50-100s in the plot.



>
>Thank you,
>
>Andrea
>
>-----Original Message-----
>From: Rong Pan (ropan) [mailto:[e-mail]]
>Sent: Thursday, August 13, 2015 9:56 PM
>To: Francini, Andrea (Andrea); Roland Bless; Jonathan Morton
>Cc: [e-mail]
>Subject: Re: [aqm] PIE vs. RED
>
>>>That's my point (and I believe also Roland's): In a direct experimental
>>>comparison, this good property of PIE would be better emphasized against
>>>a >>configuration of RED where the queue length thresholds are not
>>>grossly oversized.
>
>
>In our PIE paper (attached), Figure 8 shows the max latency of RED is
>around 100ms (which is very reasonable). PIE controls latency regardless
>of traffic intensity. It is the plot that you want.
>
>Thanks,
>
>Rong
>
>On 8/13/15, 5:25 PM, "Francini, Andrea (Andrea)"
><[e-mail]> wrote:
>
>>> Delayed-based RED still would associate latency with drop probability:
>>> drop probability will only go up when queueing latency goes up. A
>>>higher
>>> drop probability can only be achieved via higher queueing latency.  As
>>>we
>>> proved in PIE, the two can be made independent. We can maintain low
>>> latency regardless of traffic intensity.
>>> 
>>> Rong
>>
>>That's my point (and I believe also Roland's): In a direct experimental
>>comparison, this good property of PIE would be better emphasized against
>>a configuration of RED where the queue length thresholds are not grossly
>>oversized.
>>
>>By the way, Global Synchronization Protection (GSP) also drops/marks at a
>>fixed delay level independently of the drop/mark rate that keeps the
>>queue stable. The draft that describes it
>>(https://tools.ietf.org/id/draft-lauten-aqm-gsp-02.txt) is still active.
>>I have seen only marginal comments about GSP. Any specific reason why?
>>
>>Andrea 
>>
>>
>>On 8/13/15, 4:07 PM, "aqm on behalf of Francini, Andrea (Andrea)"
>><[e-mail] on behalf of [e-mail]>
>>wrote:
>>
>>>To second Roland's point, the advantage of PIE over RED should not be
>>>entirely in the use of delay-based thresholds instead of queue-length
>>>ones, otherwise it could be argued that a version of RED with
>>>delay-based
>>>thresholds is not too hard to design (Wolfram easily did it for his GSP
>>>scheme). 
>>>
>>>With such a RED version in place, hopefully PIE would still show better
>>>performance, so the same superiority should also emerge when the
>>>queue-length thresholds of conventional RED are reasonably tuned around
>>>the traffic scenario of each experiment.
>>>
>>>Andrea
>>>
>>>-----Original Message-----
>>>From: aqm [mailto:[e-mail]] On Behalf Of Roland Bless
>>>Sent: Thursday, August 13, 2015 6:39 PM
>>>To: Jonathan Morton
>>>Cc: [e-mail]
>>>Subject: Re: [aqm] PIE vs. RED
>>>
>>>Hi Jonathan,
>>>
>>>Am 13.08.2015 um 19:35 schrieb Jonathan Morton:
>>>> In the real world, the hardware buffer size is rarely matched to the
>>>> real BDP.  There are several reasons for this, but a couple of
>>>> fundamental ones are:
>>>> 
>>>> - BDP varies with RTT, which is in general different for flows
>>>> simultaneously using the same link/queue to reach different remote
>>>> hosts, and therefore cannot be accurately predicted by a hardware
>>>>vendor.
>>>
>>>Yep, sure. My point was not to promote setting the buffers according to
>>>"the BDP", but rather arguing that one should use comparable target
>>>settings when comparing AQMs, see below...
>>>
>>>> - Frequently, the queue size is tuned for the maximum capability of
>>>>the
>>>> device and a pessimistic value for RTT, but the same hardware is more
>>>> often used (at least initially) at lower link speeds and thebqueue
>>>>size
>>>> is not adjusted to compensate.  Eg. DOCSIS 2 cable but DOCSIS 3 modem,
>>>> Ethernet NIC or switch capable of 1000Mbps but operating at 100 or
>>>>even
>>>> 10, 802.11ac wifi struggling with a marginal 802.11g link...
>>>> 
>>>> Thus substantially oversized raw buffers are quite normal.  It is
>>>>AQM's
>>>> job to keep the *actual* queue occupancy low; with a properly
>>>> functioning AQM, the effects of an oversized raw queue are nil.
>>>
>>>That's correct. However, IMHO if one compares AQMs one should set/tune
>>>the individual parameters of the AQMs so that they achieve a similar
>>>target value (and not more than an order of magnitude apart).
>>>This is probably relevant for the aqm eval guidelines, but
>>>I'll come up with a detailed review for the draft within the next
>>>days...
>>>
>>>Regards,
>>> Roland
>>>
>>>_______________________________________________
>>>aqm mailing list
>>>[e-mail]
>>>https://www.ietf.org/mailman/listinfo/aqm
>>>
>>>_______________________________________________
>>>aqm mailing list
>>>[e-mail]
>>>https://www.ietf.org/mailman/listinfo/aqm
>>
>