Congestion Management for Data Centers: IEEE 802.1 Ethernet Standard Balaji Prabhakar Departments of EE and CS Stanford University

Background Data Centers see the true convergence of L3 and L2 transport While TCP is the dominant L3 transport protocol, and a significant amount of L2 traffic uses it, there is other L2 traffic; notably, storage and media This, and other reasons, have prompted the IEEE 802.1 standards body to develop an Ethernet congestion management standard In this lecture, we shall see the development of the QCN (Quantized Congestion Notification) algorithm for standardization in the IEEE 802.1 Data Center Bridging standards We will also review the technical background of congestion control research The lecture has 3 parts A brief overview of the relevant congestion control background A description of the QCN algorithm and its performance The Averaging Principle: A new control-theoretic idea underlying the QCN and BIC-TCP algorithms which stabilizes them when loop delays increase; very useful for operating high- speed links with shallow buffers---the situation in 10+ Gbps Ethernets

Managing Congestion Congestion is a standard feature of networked systems; in data networks, Congestion occurs when links are oversubscribed when traffic and/or link bandwidth changes A congestion notification mechanism allows switches/routers to directly control the rate of the ultimate sources of the traffic We’ve been involved in developing QCN (for Quantized Congestion Notification) for standardization in the Data Center Bridging track of the IEEE 802.1 Ethernet standards For deployment in 10 (and 40 and 100) Gbps Data Center Ethernets Complete information on the QCN algorithm (p-code, draft of standard, detailed simulations of lots of scenarios) available at

Congestion control in the Internet Queue management schemes (e.g. RED) at the links signal congestion by either dropping or marking packets using ECN TCP at end-systems uses these signals to vary the sending rate There exists a rich history of algorithm development, control-theoretic analysis and detailed simulation of queue management schemes and congestion control algorithms for the Internet Jacobson, Floyd et al, Kelly et al, Low et al, Srikant et al, Misra et al, Katabi et al … TCP is excellent, so why look for another algorithm? There is other traffic on Ethernet than TCP; so, native Ethernet congestion management is needed TCP’s “one size fits all” approach makes it too conservative for high bandwidth-delay product networks A hardware-based algorithm is needed for the very high speeds of operation encountered in 10, 40 and 100 Gbps Ethernet and the Internet have very different operating conditions

Switched Ethernet vs. the Internet Some significant differences … No per-packet acks in Ethernet, unlike in the Internet Not possible to know round trip time! So congestion must be signaled to the source by switches Algorithm not automatically self-clocked (like TCP) Links can be paused; i.e. packets may not be dropped No sequence numbering of L2 packets Sources do not start transmission gently (like TCP slow-start); they can potentially come on at the full line rate of 10Gbps Ethernet switch buffers are much smaller than router buffers (100s of KBs vs 100s of MBs) Most importantly, algorithm should be simple enough to be implemented completely in hardware Note: The QCN algorithm we have developed has Internet relatives; notably BIC-TCP at the source and the REM/PI controllers at switches

L2 Transport: IEEE 802.1 IEEE 802.1 Data Center Bridging standards: Enhancements to Ethernet Reliable delivery (802.1Qbb): Link-level flow control (PAUSE) prevents congestion drops Ethernet congestion management (802.1Qau): Prevents congestion spreading due to PAUSE Consequences Hardware-friendly algorithms: can operate on 10—100Gbps links Partial offload of CPU: no packet retransmissions Corruption losses require abort/restart; 10G over copper uses short cables to keep low BER PAUSE absorption buffers: proportional to bdwdth x delay of links, high memory bandwidth NOTE: Recent work addresses the last two points; this is not covered in the course Pause absorption buffers X X Congestion spreading

Overview of Congestion Control Research

Stability Congestion control algorithms aim to deliver high throughput, maintain low latencies/backlogs, be fair to all flows, be simple to implement and easy to deploy Performance is related to stability of control loop “Stability” refers to the non-oscillatory or non-exploding behavior of congestion control loops. In real terms, stability refers to the non-oscillatory behavior of the queues at the switch. If the switch buffers are short, oscillating queues can overflow (hence drop packets/pause the link) or underflow (hence lose utilization) In either case, links cannot be fully utilized, throughput is lost, flow transfers take longer So stability is an important property, especially for networks with high bandwidth-delay products operating with shallow buffers

Unit step response of the network The control loops are not easy to analyze They are described by non-linear, delay differential equations which are usually impossible to analyze So linearized analyses are performed using Nyquist or Bode theory Is linearized analysis useful? Yes! It is not difficult to know if a zero-delay non-linear system is stable. As the delay increases, linearization can be used to tell if the system is stable for delay (or number of sources) in some range; i.e. we get sufficient conditions The above stability theory is essentially studying the “unit step response” of a network Apply many “infinitely long flows” at time 0 and see how long the network takes to settle them to the correct collective and individual rate; the first is about throughput, the second is about fairness

TCP--RED: A basic control loop minth maxth qavg p RED: Drop probability, p, increases as the congestion level goes up TCP: Slow start + Congestion avoidance Congestion avoidance: AIMD No loss: increase window by 1; Pkt loss: cut window by half

TCP--RED Two ways to analyze and understand this control loop Simulations: ns-2 Theory: Delay-differential equations ns-2: A widely used event-driven simulator for the Internet Very detailed and accurate Different types of transport protocols: TCP, UDP, … Router mechanisms and algorithms: RED, DRR, … Web traffic: sessions, flows, power law flow sizes, … Different types of network: wired, wireless, satellite, mobility,…

The simulation setup grp2 grp1 RED grp3 100Mbps # of TCP flows 1200 50 150 200 100 time 300 # of TCP flows 1200 50 150 200 100 time 100Mbps grp1 grp2 grp3 RED # of TCP flows 600 50 150 200 100 time

Delay at Link 1

TCP--RED: Analytical model RED Control Time Delay - 1/R p C TCP Control q

TCP--RED: Analytical model Users: 1.5 Network: W: window size; RTT: round trip time; C: link capacity q: queue length; qa: ave queue length p: drop probability *By V. Misra, W. Dong and D. Towsley at SIGCOMM 2000 *Fluid model concept originated by F. Kelly, A. Maullo and D. Tan at Jour. Oper. Res. Society, 1998

Accuracy of analytical model Recall the ns-2 simulation from earlier: Delay at Link 1

Accuracy of analytical model

Accuracy of analytical model

Why are the Diff Eqn models so accurate? They’ve been developed in Physics, where they are called Mean Field Models The main idea very difficult to model large-scale systems: there are simply too many events, too many random quantities but, it is quite easy to model the mean or average behavior of such systems interestingly, when the size of the system grows, its behavior gets closer and closer to that predicted by the mean-field model! physicists have been exploiting this feature to model large magnetic materials, gases, etc. just as a few electrons/particles don’t have a very big influence on a system, so is Internet resource usage not heavily influenced by a few packets: aggregates matter more

TCP--RED: Stability analysis Given the differential equations, in principle one can figure out whether the TCP--RED control loop is stable However, the differential equations are very complicated 3rd or 4th order, nonlinear, with delays There is no general theory, specific case treatments exist “Linearize and analyze” Linearize equations around the (unique) operating point Analyze resultant linear, delay-differential equations using Nyquist or Bode theory End result: Design stable control loops Determine stability conditions (RTT limits, number of users, etc) Obtain control loop parameters: gains, drop functions, …

Instability of TCP--RED As the bandwidth-delay-product increases, the TCP--RED control loop becomes unstable Parameters: 50 sources, link capacity = 9000 pkts/sec, TCP--RED Source: S. Low et. al. Infocom 2002

Summary We saw a very brief overview of research on the analysis of congestion control systems As loop lags increase, the control loop becomes very oscillatory This is true of any control scheme, not just congestion control schemes In networks, oscillatory queue sizes tend to underflow buffers, causing to a loss of throughput; especially true for high BDP networks with shallow buffers This has led to much research on developing algorithms for high BDP networks; e.g. High-Speed TCP, XCP, RCP, Scalable TCP, BIC-TCP, etc We shall return to this later, after describing the QCN algorithm we have developed for the IEEE 802.1 standard

Ashvin Lakshmikantha, Broadcom Quantized Congestion Notification (QCN): Congestion control for Ethernet Joint work with: Mohammad Alizadeh, Berk Atikoglu and Abdul Kabbani, Stanford University Ashvin Lakshmikantha, Broadcom Rong Pan, Cisco Systems Mick Seaman, Chair, Security Group; Ex-Chair, Interworking Group, IEEE 802.1

Overview The description of QCN is brief, restricted to the main points of the algorithm A fuller description is available at the IEEE 802.1 Data Center Bridging Task Group’s website, including extensive simulations and pseudo-code We will describe the congestion control loop How is congestion measured at the switches? What is the signal? And, how does the switch send it? (Remember there are no per-packet acks in Ethernet) What does the source do when it receives a congestion signal? Terminology: Congestion Point: Where congestion occurs, mainly switches Reaction Point: Source of traffic, mainly rate limiters in Ethernet NICs

QCN: Congestion Point Dynamics Consider the single-source, single-switch loop below Congestion Point (Switch) Dynamics: Sample packets, compute feedback (Fb), quantize Fb to 6 bits, and reflect only negative Fb values back to Reaction Point with a probability proportional to Fb. Qeq Source Pmax Reflection Probability Fb = -(Q-Qeq + w . dQ/dt ) = -(queue offset + w.rate offset) Pmin |Fb|

Congestion message recd QCN: Reaction Point Source (reaction point): Transmit regular Ethernet frames. When congestion message arrives: Multiplicative Decrease: Fast Recovery similar to BIC-TCP: gives high performance in high bandwidth- delay product networks, while being very simple. Active Probing Fast Recovery Time Rate Current Rate Congestion message recd Rd Rd/2 Rd/4 Rd/8 Target Rate CR TR Active Probing

Timer-supported QCN Byte-Ctr RL Timer Byte-Counter RL Timer 5 cycles of FR (150KB per cycle) AI cycles afterwards (75KB per cycle) Fb < 0 sends timer to FR Byte-Ctr RL In FR if both byte-ctr and timer in FR In AI if only one of byte-ctr or timer in AI In HAI if both byte-ctr and timer in AI Note: RL goes to HAI only after 500 pkts have been sent RL Timer Timer 5 cycles of FR (T msec per cycle) AI cycles afterwards (T/2 msec/cycle) Fb < 0 sends timer to FR

Simulations: Basic Case Parameters 10 sources share a 10 G link, whose capacity drops to 0.5G during 2-4 secs Max offered rate per source: 1.05G RTT = 50 usec Buffer size = 100 pkts (150KB); Qeq = 22 T = 10 msecs RAI = 5 Mbps RHAI = 50 Mbps 10 G 10 G Source 1 Source 2 0.5G Source 10

Recovery Time Recovery time = 80 msec

Fluid Model for QCN P = Φ(Fb) Assume N flows pass through a single queue at a switch. State variables are TRi(t), CRi(t), q(t), p(t). 10% Fb 63

Accuracy: Equations vs. ns-2 simulations N = 10, RTT = 100 us N = 100, RTT = 500 us N = 10, RTT = 1 ms N = 10, RTT = 2 ms

Summary The algorithm has been extensively tested in deployment scenarios of interest Esp. interoperability with link-level PAUSE and TCP All presentations are available at the IEEE 802.1 website: The theoretical development is interesting, but most notably because QCN (and BIC-TCP) display strong stability in the face of increasing lags, or, equivalently in high bandwidth-delay product networks While attempting to understand why these schemes perform so well, we have uncovered a method for improving the stability of any congestion control scheme; we present this next

The Averaging Principle

Background to the AP When the lags in a control loop increase, the system becomes oscillatory and eventually becomes unstable Feedback compensation is applied to restore stability; the two main flavors of feedback compensation in are: Determine lags (round trip times), apply the correct “gains” for the loop to be stable (e.g. XCP, RCP, FAST). Include higher order queue derivatives in the congestion information fed back to the source (e.g. REM/PI, BCN). Method 1 is not suitable for us, we don’t know RTTs in Ethernet Method 2 requires a change to the switch implementation The Averaging Principle is a different method It is suited to Ethernet where round trip times are unavailable It doesn’t need more feedback, hence switch implementations don’t have to change QCN and BIC-TCP already turn out to employ it

The Averaging Principle (AP)‏ A source in a congestion control loop is instructed by the network to decrease or increase its sending rate (randomly) periodically AP: a source obeys the network whenever instructed to change rate, and then voluntarily performs averaging as below TR = Target Rate CR = Current Rate

Recall: QCN does 5 steps of Averaging The Fast Recovery portion of QCN, there are 5 steps of averaging In fact, QCN and BIC-TCP are the Ave Prin applied to TCP! Time Rate Current Rate Congestion message recd Rd Rd/2 Rd/4 Rd/8 Target Rate CR TR Active Probing

Applying the AP RCP: Rate Control Protocol Dukkipatti and McKeown A router computes an upper bound R on the rate of all flows traversing it. R recomputed every T (= 10) msec as follows: ‏ Where d0: Round trip time estimate (set constant= 10 msec in our case)‏ C: link capacity (= 2.4 Gbps) Q: Current queue size at the switch y(t): incoming rate α = 0.1 ß = 1 A flow chooses the smallest advertised rate on its path. We consider a scenario where 10 RCP sources share a single link.

AP-RCP Stability RTT = 60 msec RTT = 65 msec

AP-RCP Stability cont’d RTT = 120 msec RTT = 130 msec

AP-RCP Stability cont’d RTT = 230 msec RTT = 240 msec

Understanding the AP As mentioned earlier, the two major flavors of feedback compensation are: Determine lags, chose appropriate gains Feedback higher derivatives of state We prove that the AP is sense equivalent to both of the above! This is great because we don’t need to change network routers and switches And the AP is really very easy to apply; no lag-dependent optimizations of gain parameters needed

AP Equivalence: Single Source Case does AP Fb Regular source 0.5 Fb + 0.25 T dFb/dt Systems 1 and 2 are discrete-time models for an AP enabled source, and a regular source respectively. Main Result: Systems 1 and 2 are algebraically equivalent. That is, given identical input sequences, they produce identical output sequences. Therefore the AP is equivalent to adding a derivative to the feedback and reducing the gain! Thus, the AP does both known forms of feedback compensation without knowing RTTs or changing switch implementations

AP-RCP vs PD-RCP RTT = 120 msec RTT = 130 msec

A Generic Control Example As an example, we consider the plant transfer function: P(s) = (s+1)/(s3+1.6s2+0.8s+0.6)

Step Response Basic AP, No Delay

Step Response Basic AP, Delay = 8 seconds

Step Response Two-step AP, Delay = 14 seconds

Step Response Two-step AP, Delay = 25 seconds Two-step AP is even more stable than Basic AP

Summary of AP The AP is a simple method for making many control loops (not just congestion control loops) more robust to increasing lags Gives a clear understanding as to the reason why the BIC-TCP and QCN algorithms have such good delay tolerance: they do averaging repeatedly There is a theorem which deals explicitly with the QCN-type loop Variations of the basic principle are possible; i.e. average more than once, average by more than half-way, etc The theory is fairly complete in these cases

QCN and Buffer Sizing

Background: TCP Buffer Sizing Standard “rule of thumb”: Single TCP flow: Bandwidth × Delay worth of buffering needed for 100 % utilization. Recent result (Appenzellar et al.): For N >> 1 TCP flows: Bdwdth x Delay/sqrt(N) amount of buffering is enough. The essence of this result is that when many flows combine, the Variance of the net sending rate decreases: Buffer sizing problem is challenging in data centers: Typically, only a small number of flows are active on each path. (N is small) Ethernet switches are typically built with shallow buffers to keep costs down.

Example: Simulation Setup Switch 10 Gig Ethernet Switch buffer is 150 Kbytes deep. We compare TCP and QCN for various # of flows, and RTTs.

TCP vs QCN (N = 1, RTT = 120 μs) TCP QCN Throughput = 99.5% Standard Deviation = 265.4 Mbps Throughput = 99.5% Standard Deviation = 13.8 Mbps

TCP vs QCN (N = 1, RTT = 250 μs) TCP QCN Throughput = 95.5% Standard Deviation = 782.7 Mbps Throughput = 99.5% Standard Deviation = 33.3 Mbps

TCP vs QCN (N = 1, RTT = 500 μs) TCP QCN Throughput = 88% Standard Deviation = 1249.7 Mbps Throughput = 99.5% Standard Deviation = 95.4 Mbps

TCP vs QCN (N = 10, RTT = 120 μs) TCP QCN Throughput = 99.5% Standard Deviation = 625.1 Mbps Throughput = 99.5% Standard Deviation = 25.1 Mbps

TCP vs QCN (N = 10, RTT = 250 μs) TCP QCN Throughput = 95.5% Standard Deviation = 981 Mbps Throughput = 99.5% Standard Deviation = 27.2 Mbps

TCP vs QCN (N = 10, RTT = 500 μs) TCP QCN Throughput = 89% Standard Deviation = 1311.4 Mbps Throughput = 99.5% Standard Deviation = 170.5 Mbps

QCN and shallow buffers In contrast to TCP, QCN is stable with shallow buffers, even with few sources. Why? Recall that buffer requirements are closely related to sending rate variance: Buffer size = C x Var(R1) x Bdwdth x Delay/ sqrt(N) TCP: Good performance for large N, since the denominator is large. QCN: Good performance for all N, since the numerator is small. Thus, averaging reduces the variance of a source’s sending rate This is a stochastic interpretation of the Averaging Principle’s success in keeping stability with shallow buffers

Conclusions We have seen the background, development and analysis of a congestion control scheme for the IEEE 802.1 Ethernet standard The QCN algorithm is More stable with respect to control loop delays Requires much smaller buffers than TCP Easy to build in hardware The Averaging Principle is interesting and we’re exploring its use in nonlinear control systems