Internet Draft                                                    R. Pan
Network Working Group                 P. Natarajan, F. Baker, B.V. Steeg
Intended Status: Informational    M. Prabhu, V. Subramanian, C. Piglione
                                                           Cisco Systems

Expires: March 21, 2015                               September 17, 2014

           PIE: A Lightweight Control Scheme To Address the
                          Bufferbloat Problem



   Bufferbloat is a phenomenon where excess buffers in the network cause
   high latency and jitter. As more and more interactive applications
   (e.g. voice over IP, real time video streaming and financial
   transactions) run in the Internet, high latency and jitter degrade
   application performance. There is a pressing need to design
   intelligent queue management schemes that can control latency and
   jitter; and hence provide desirable quality of service to users.

   We present here a lightweight design, PIE(Proportional Integral
   controller Enhanced) that can effectively control the average
   queueing latency to a target value. Simulation results, theoretical
   analysis and Linux testbed results have shown that PIE can ensure low
   latency and achieve high link utilization under various congestion
   situations. The design does not require per-packet timestamp, so it
   incurs very small overhead and is simple enough to implement in both
   hardware and software.

Status of this Memo

   This Internet-Draft is submitted to IETF in full conformance with the
   provisions of BCP 78 and BCP 79.

   Internet-Drafts are working documents of the Internet Engineering
   Task Force (IETF), its areas, and its working groups.  Note that
   other groups may also distribute working documents as

   Internet-Drafts are draft documents valid for a maximum of six months
   and may be updated, replaced, or obsoleted by other documents at any
   time.  It is inappropriate to use Internet-Drafts as reference
   material or to cite them other than as "work in progress."

Pan et al.               Expires March 21, 2015                 [Page 1]

INTERNET DRAFT                    PIE                     March 21, 2015

   The list of current Internet-Drafts can be accessed at

   The list of Internet-Draft Shadow Directories can be accessed at

Copyright and License Notice

   Copyright (c) 2012 IETF Trust and the persons identified as the
   document authors. All rights reserved.

   This document is subject to BCP 78 and the IETF Trust's Legal
   Provisions Relating to IETF Documents
   ( in effect on the date of
   publication of this document. Please review these documents
   carefully, as they describe your rights and restrictions with respect
   to this document. Code Components extracted from this document must
   include Simplified BSD License text as described in Section 4.e of
   the Trust Legal Provisions and are provided without warranty as
   described in the Simplified BSD License.

Table of Contents

   1. Introduction  . . . . . . . . . . . . . . . . . . . . . . . . .  3
   2. Terminology . . . . . . . . . . . . . . . . . . . . . . . . . .  4
   3. Design Goals  . . . . . . . . . . . . . . . . . . . . . . . . .  4
   4. The PIE Scheme  . . . . . . . . . . . . . . . . . . . . . . . .  5
     4.1 Random Dropping  . . . . . . . . . . . . . . . . . . . . . .  5
     4.2 Drop Probability Calculation . . . . . . . . . . . . . . . .  6
     4.3 Departure Rate Estimation  . . . . . . . . . . . . . . . . .  7
     4.4 Handling Bursts  . . . . . . . . . . . . . . . . . . . . . .  8
   5. Comments and Discussions  . . . . . . . . . . . . . . . . . . .  9
   8. IANA Considerations . . . . . . . . . . . . . . . . . . . . . . 11
   9. References  . . . . . . . . . . . . . . . . . . . . . . . . . . 11
     9.1  Normative References  . . . . . . . . . . . . . . . . . . . 11
     9.2  Informative References  . . . . . . . . . . . . . . . . . . 11
     9.3  Other References  . . . . . . . . . . . . . . . . . . . . . 11
   Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . . 12

Pan et al.               Expires March 21, 2015                 [Page 2]

INTERNET DRAFT                    PIE                     March 21, 2015

1. Introduction

   The explosion of smart phones, tablets and video traffic in the
   Internet brings about a unique set of challenges for congestion
   control. To avoid packet drops, many service providers or data center
   operators require vendors to put in as much buffer as possible. With
   rapid decrease in memory chip prices, these requests are easily
   accommodated to keep customers happy. However, the above solution of
   large buffer fails to take into account the nature of the TCP
   protocol, the dominant transport protocol running in the Internet.
   The TCP protocol continuously increases its sending rate and causes
   network buffers to fill up. TCP cuts its rate only when it receives a
   packet drop or mark that is interpreted as a congestion signal.
   However, drops and marks usually occur when network buffers are full
   or almost full. As a result, excess buffers, initially designed to
   avoid packet drops, would lead to highly elevated queueing latency
   and jitter. It is a delicate balancing act to design a queue
   management scheme that not only allows short-term burst to smoothly
   pass, but also controls the average latency when long-term congestion

   Active queue management (AQM) schemes, such as Random Early Discard
   (RED), have been around for well over a decade. AQM schemes could
   potentially solve the aforementioned problem. RFC 2309[RFC2309]
   strongly recommends the adoption of AQM schemes in the network to
   improve the performance of the Internet. RED is implemented in a wide
   variety of network devices, both in hardware and software.
   Unfortunately, due to the fact that RED needs careful tuning of its
   parameters for various network conditions, most network operators
   don't turn RED on. In addition, RED is designed to control the queue
   length which would affect delay implicitly. It does not control
   latency directly. Hence, the Internet today still lacks an effective
   design that can control buffer latency to improve the quality of
   experience to latency-sensitive applications.

   Recently, a new AQM scheme, CoDel[CoDel],  was proposed to control
   the latency directly to address the bufferbloat problem. CoDel
   requires per packet timestamps. Also, packets are dropped at the
   dequeue function after they have been enqueued for a while. Both of
   these requirements consume excessive processing and infrastructure
   resources. This consumption will make CoDel expensive to implement
   and operate, especially in hardware.

   PIE aims to combine the benefits of both RED and CoDel: easy to
   implement like RED and directly control latency like CoDel. Similar
   to RED, PIE randomly drops a packet at the onset of the congestion.
   The congestion detection, however, is based on the queueing latency
   like CoDel instead of the queue length like RED. Furthermore, PIE

Pan et al.               Expires March 21, 2015                 [Page 3]

INTERNET DRAFT                    PIE                     March 21, 2015

   also uses the latency moving trends: latency increasing or
   decreasing, to help determine congestion levels. The design
   parameters of PIE are chosen via stability analysis. While these
   parameters can be fixed to work in various traffic conditions, they
   could be made self-tuning to optimize system performance.

   In addition, we assume any delay-based AQM scheme would be applied
   over a Fair Queueing (FQ) structure or its approximate design, Class
   Based Queueing (CBQ). FQ is one of the most studied scheduling
   algorithms since it was first proposed in 1985 [RFC970]. CBQ has been
   a standard feature in most network devices today[CBQ]. These designs
   help flows/classes achieve max-min fairness and help mitigate bias
   against long flows with long round trip times(RTT). Any AQM scheme
   that is built on top of FQ or CBQ could benefit from these
   advantages. Furthermore, we believe that these advantages such as per
   flow/class fairness are orthogonal to the AQM design whose primary
   goal is to control latency for a given queue. For flows that are
   classified into the same class and put into the same queue, we need
   to ensure their latency is better controlled and their fairness is
   not worse than those under the standard DropTail or RED design.

   This draft describes the overall design goals, system elements and
   implementation details of PIE. We will also discuss various design
   considerations, including how auto-tuning can be done.

2. Terminology

   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   document are to be interpreted as described in RFC 2119 [RFC2119].

3. Design Goals

   We explore a queue management framework where we aim to improve the
   performance of interactive and delay-sensitive applications. The
   design of our scheme follows a few basic criteria.

        * First, we directly control queueing latency instead of
        controlling queue length. Queue sizes change with queue draining
        rates and various flows' round trip times. Delay bloat is the
        real issue that we need to address as it impairs real time
        applications. If latency can be controlled, bufferbloat is not
        an issue. As a matter of fact, we would allow more buffers for
        sporadic bursts as long as the latency is under control.

Pan et al.               Expires March 21, 2015                 [Page 4]

INTERNET DRAFT                    PIE                     March 21, 2015

        * Secondly, we aim to attain high link utilization. The goal of
        low latency shall be achieved without suffering link under-
        utilization or losing network efficiency. An early congestion
        signal could cause TCP to back off and avoid queue building up.
        On the other hand, however, TCP's rate reduction could result in
        link under-utilization. There is a delicate balance between
        achieving high link utilization and low latency.

        * Furthermore, the scheme should be simple to implement and
        easily scalable in both hardware and software. The wide adoption
        of RED over a variety of network devices is a testament to the
        power of simple random early dropping/marking. We strive to
        maintain similar design simplicity.

        * Finally, the scheme should ensure system stability for various
        network topologies and scale well with arbitrary number streams.
        Design parameters shall be set automatically. Users only need to
        set performance-related parameters such as target queue delay,
        not design parameters.

In the following, we will elaborate on the design of PIE and its

4. The PIE Scheme

As illustrated in Fig. 1, our scheme comprises three simple components:
a) random dropping at enqueing; b) periodic drop probability update; c)
dequeing rate estimation.

The following sections describe these components in further detail, and
explain how they interact with each other.  At the end of this section,
we will discuss how the scheme can be easily augmented to precisely
control bursts.

4.1 Random Dropping

Like any state-of-the-art AQM scheme, PIE would drop packets randomly
according to a drop probability, p, that is obtained from the drop-
probability-calculation component:

    * upon a packet arrival

        randomly drop a packet with a probability p.

Pan et al.               Expires March 21, 2015                 [Page 5]

INTERNET DRAFT                    PIE                     March 21, 2015

         Random Drop
             /               --------------
     -------/  -------------->    | | | | | -------------->
            /|\                   | | | | |         |
             |               --------------         |
             |                 Queue Buffer         |
             |                     |                | Departure bytes
             |                     |queue           |
             |                     |length          |
             |                     |                |
             |                    \|/              \|/
             |          -----------------    -------------------
             |          |     Drop      |    |                 |
             -----<-----|  Probability  |<---| Departure Rate  |
                        |  Calculation  |    | Estimation      |
                        -----------------    -------------------

                      Figure 1. The PIE Structure

4.2 Drop Probability Calculation

The PIE algorithm periodically updates the drop probability as follows:

    * estimate current queueing delay using Little's law:

        est_del = qlen/depart_rate;

    * calculate drop probability p as:

        p = p + alpha*(est_del-target_del) + beta*(est_del-est_del_old);

        est_del_old = est_del.

Here, the current queue length is denoted by qlen. The draining rate of
the queue, depart_rate, is obtained from the departure-rate-estimation
block.  Variables, est_del and est_del_old, represent the current and
previous estimation of the queueing delay. The target latency value is
expressed in target_del.  The update interval is denoted as Tupdate.

Note that the calculation of drop probability is based not only on the
current estimation of the queueing delay, but also on the direction

Pan et al.               Expires March 21, 2015                 [Page 6]

INTERNET DRAFT                    PIE                     March 21, 2015

where the delay is moving, i.e., whether the delay is getting longer or
shorter. This direction can simply be measured as the difference between
est_del and est_del_old. This is the classic Proportional Integral
controller design that is adopted here for controlling queueing latency.
The controller parameters, in the unit of hz, are designed using
feedback loop analysis where TCP's behaviors are modeled using the
results from well-studied prior art[TCP-Models].

We would like to point out that this type of controller has been studied
before for controlling the queue length [PI, QCN]. PIE adopts the
Proportional Integral controller for controlling delay and makes the
scheme auto-tuning. The theoretical analysis of PIE is under paper
submission and its reference will be included in this draft once it
becomes available. Nonetheless, we will discuss the intuitions for these
parameters in Section 5.

4.3 Departure Rate Estimation

The draining rate of a queue in the network often varies either because
other queues are sharing the same link, or the link capacity fluctuates.
Rate fluctuation is particularly common in wireless networks. Hence, we
decide to measure the departure rate directly as follows.

    * we are in a measurement cycle if we have enough data in the queue:

        qlen > deq_threshold

    * if in a measurement cycle:

        upon a packet departure

        dq_count = dq_count + deque_pkt_size;

    * if dq_count > deq_threshold then

        depart_rate = dq_count/(now-start);

        dq_count = 0;

        start = now;

We only measure the departure rate when there are sufficient data in the

Pan et al.               Expires March 21, 2015                 [Page 7]

INTERNET DRAFT                    PIE                     March 21, 2015

buffer, i.e., when the queue length is over a certain threshold,
deq_threshold. Short, non-persistent bursts of packets result in empty
queues from time to time, this would make the measurement less accurate.
The parameter, dq_count, represents the number of bytes departed since
the last measurement. Once dq_count is over a certain threshold,
deq_threshold, we obtain a measurement sample. The threshold is
recommended to be set to 10KB assuming a typical packet size of around
1KB or 1.5KB. This threshold would allow us a long enough period to
obtain an average draining rate but also fast enough to reflect sudden
changes in the draining rate. Note that this threshold is not crucial
for the system's stability.

4.4 Handling Bursts

The above three components form the basis of the PIE algorithm. Although
we aim to control the average latency of a congested queue, the scheme
should allow short term bursts to pass through the system without
hurting them. We would like to discuss how PIE manages bursts in this

Bursts are well tolerated in the basic scheme for the following reasons:
first, the drop probability is updated periodically. Any short term
burst that occurs within this period could pass through without
incurring extra drops as it would not trigger a new drop probability
calculation. Secondly, PIE's drop probability calculation is done
incrementally. A single update would only lead to a small incremental
change in the probability. So if it happens that a burst does occur at
the exact instant that the probability is being calculated, the
incremental nature of the calculation would ensure its impact is kept

Nonetheless, we would like to give users a precise control of the burst.
We introduce a parameter, max_burst, that is similar to the burst
tolerance in the token bucket design. By default, the parameter is set
to be 100ms. Users can certainly modify it according to their
application scenarios. The burst allowance is added into the basic PIE
design as follows:

    * if p == 0 and est_del < del_ref and est_del_old < del_ref

        burst_allowance = max_burst;

    * upon packet arrival

        if burst_allowance > 0 enqueue packet;

    * upon probability update

Pan et al.               Expires March 21, 2015                 [Page 8]

INTERNET DRAFT                    PIE                     March 21, 2015

        burst_allowance = burst_allowance - Tupdate;

The burst allowance, noted by burst_allowance, is initialized to
max_burst. As long as burst_allowance is above zero, an incoming packet
will be enqueued bypassing the random drop process. During each update
instance, the value of burst_allowance is decremented by the update
period, Tupdate. When the congestion goes away, defined by us as p
equals to 0 and both the current and previous samples of estimated delay
are less than target_del, we reset burst_allowance to max_burst.

5. Comments and Discussions

While the formal analysis will be included later, we would like to
discuss the intuitions regarding how to determine the key parameters.
Although the PIE algorithm would set them automatically, they are not
meant to be magic numbers. We hope to give enough explanations here to
help demystify them so that users can experiment and explore on their

As it is obvious from the above, the crucial equation in the PIE
algorithm is

     p = p + alpha*(est_del-target_del) + beta*(est_del-est_del_old).

The value of alpha determines how the deviation of current latency from
the target value affects the drop probability.  The beta term exerts
additional adjustments depending on whether the latency is trending up
or down. Note that the drop probability is reached incrementally, not
through a single step. To avoid big swings in adjustments which often
leads to instability, we would like to tune p in small increments.
Suppose that p is in the range of 1%. Then we would want the value of
alpha and beta to be small enough, say 0.1%, adjustment in each step. If
p is in the higher range, say above 10%, then the situation would
warrant a higher single step tuning, for example 1%. Finally, the drop
probability would only be stabilized when the latency is stable, i.e.
est_del equals est_del_old; and the value of the latency is equal to
target_del. The relative weight between alpha and beta determines the
final balance between latency offset and latency jitter.

The update interval, Tupdate, also plays a key role in stability. Given
the same alpha and beta values, the faster the update is, the higher the
loop gain will be. As it is not showing explicitly in the above
equation, it can become an oversight. Notice also that alpha and beta
have a unit of hz.

As a further extension, we could introduce weights for flows that are

Pan et al.               Expires March 21, 2015                 [Page 9]

INTERNET DRAFT                    PIE                     March 21, 2015

classified into the same queue to achieve differential dropping. For
example, the dropping probability for flow i could be p(i) =
p/weight(i). Flows with higher weights would receive proportionally less
drops; and vice versa. Adding FQ on top, FQ_PIE, is another alternative.

Also, we have discussed congestion notification via the form of packet
drops. The algorithm can be easily applied to networks codes where Early
Congestion Notification (ECN) is enabled. The drop probability, p, above
would become marking probability.

6. Implementation

PIE can be applied to existing hardware or software solutions. In this
section, we discuss the implementation cost of the PIE algorithm. There
are three steps involved in PIE as discussed in Section 4. We examine
their complexities as follows.

Upon packet arrival, the algorithm simply drops a packet randomly based
on the drop probability p. This step is straightforward and requires no
packet header examination and manipulation. Besides, since no per packet
overhead, such as a timestamp, is required, there is no extra memory
requirement. Furthermore, the input side of a queue is typically under
software control while the output side of a queue is hardware based.
Hence, a drop at enqueueing can be readily retrofitted into existing
hardware or software implementations.

The drop probability calculation is done in the background and it occurs
every Tudpate interval. Given modern high speed links, this period
translates into once every tens, hundreds or even thousands of packets.
Hence the calculation occurs at a much slower time scale than packet
processing time, at least an order of magnitude slower. The calculation
of drop probability involves multiplications using alpha and beta. Since
the algorithm is not sensitive to the precise values of alpha and beta,
we can choose the values, e.g. alpha= 0.25 and beta= 2.5 so that
multiplications can be done using simple adds and shifts. As no
complicated functions are required, PIE can be easily implemented in
both hardware and software. The state requirement is only two variables
per queue: est_del and est_del_old. Hence the memory overhead is small.

In the departure rate estimation, PIE uses a counter to keep track of
the number of bytes departed for the current interval. This counter is
incremented per packet departure. Every Tupdate, PIE calculates latency
using the departure rate, which can be implemented using a
multiplication. Note that many network devices keep track an interface's
departure rate. In this case, PIE might be able to reuse this
information, simply skip the third step of the algorithm and hence
incurs no extra cost. We also understand that in some software

Pan et al.               Expires March 21, 2015                [Page 10]

INTERNET DRAFT                    PIE                     March 21, 2015

implementations, time-stamped are added for other purposes. In this
case, we can also make use of the time-stamps and bypass the departure
rate estimation and directly used the timestamp information in the drop
probability calculation.

In summary, the state requirement for PIE is limited and computation
overheads are small. Hence, PIE is simple to be implemented. In
addition, since PIE does not require any user configuration, it does not
impose any new cost on existing network management system solutions. SFQ
can be combined with PIE to provide further improvement of latency for
various flows with different priorities. However, SFQ requires extra
queueing and scheduling structures. Whether the performance gain can
justify the design overhead needs to be further investigated.

7. Incremental Deployment

One nice property of the AQM design is that it can be independently
designed and operated without the requirement of being inter-operable.

Although all network nodes can not be changed altogether to adopt
latency-based AQM schemes, we envision a gradual adoption which would
eventually lead to end-to-end low latency service for real time

8. IANA Considerations

There are no actions for IANA.

9. References

9.1  Normative References

   [RFC2119]  Bradner, S., "Key words for use in RFCs to Indicate
              Requirement Levels", BCP 14, RFC 2119, March 1997.

9.2  Informative References

   [RFC970]   Nagle, J., "On Packet Switches With Infinite
              Storage",RFC970, December 1985.

9.3  Other References

   [CoDel]      Nichols, K., Jacobson, V., "Controlling Queue Delay", ACM
                  Queue. ACM Publishing. doi:10.1145/2209249.22W.09264.

Pan et al.               Expires March 21, 2015                [Page 11]

INTERNET DRAFT                    PIE                     March 21, 2015

   [CBQ]        Cisco White Paper, "

   [TCP-Models]         Misra, V., Gong, W., and Towsley, D., "Fluid-based
                  Analysis of a Network of AQM Routers Supporting TCP
                  Flows with an Application to RED", SIGCOMM 2000.

   [PI]         Hollot, C.V., Misra, V., Towsley, D. and Gong, W.,
                "On Designing Improved Controller for AQM Routers
                Supporting TCP Flows", Infocom 2001.

   [QCN]                "Data Center Bridging - Congestion Notification",

Authors' Addresses

   Rong Pan
   Cisco Systems
   510 McCarthy Blvd,
   Milpitas, CA 95134, USA

   Preethi Natarajan,
   Cisco Systems
   510 McCarthy Blvd,
   Milpitas, CA 95134, USA

   Fred Baker
   Cisco Systems
   510 McCarthy Blvd,
   Milpitas, CA 95134, USA

   Mythili Prabhu
   Cisco Systems
   510 McCarthy Blvd,
   Milpitas, CA 95134, USA

   Chiara Piglione
   Cisco Systems
   510 McCarthy Blvd,
   Milpitas, CA 95134, USA

Pan et al.               Expires March 21, 2015                [Page 12]

INTERNET DRAFT                    PIE                     March 21, 2015

   Vijay Subramanian
   Cisco Systems
   510 McCarthy Blvd,
   Milpitas, CA 95134, USA

   Bill Ver Steeg
   Cisco Systems
   5030 Sugarloaf Parkway
   Lawrenceville, GA, 30044, USA

Pan et al.               Expires March 21, 2015                [Page 13]