1
BS HARE: Packet Queueing Delay-Driven Buffer Sharing for Datacenter Switches
arXiv:2605.24178v1 [cs.NI] 22 May 2026
Krishna Agarwal∗ , Muhamad Rizka Maulana∗ , Vamsi Addanki† , Habib Mostafaei∗ ∗ Eindhoven University of Technology, The Netherlands † Purdue University, USA
Abstract—Modern datacenter switches share packet buffers across ports to boost overall throughput and reduce packet loss. However, as buffer availability per-port-per-bandwidth unit continues to decrease, existing buffer-sharing strategies face increasing performance challenges. Recent efforts have attempted to integrate Buffer Management (BM) with Active Queue Management (AQM) to harness the advantages of both BM and AQM approaches to improve performance. While these hybrid solutions show promise, their complexity of dynamically calculating multiple factors for integration hinders generalization and efficiency. This paper presents BS HARE, a simple buffer sharing mechanism that uses packet queueing delay. BS HARE requires only a single operator-configurable parameter. Our simulation results show that BS HARE improves the flow completion time (FCT) performance of advanced transport protocols, such as PowerTCP, by up to 45.07% compared to ABM, particularly under burst-heavy datacenter workloads.
I. I NTRODUCTION Datacenter switches rely on shared on-chip buffers to absorb traffic bursts and maintain high throughput under load. However, as switch capacities have increased, buffer sizes have not kept pace, primarily due to the rising cost and design complexity of high-speed memory. As a result, the amount of buffer available per-port per unit of bandwidth has steadily decreased. This trend is particularly problematic for datacenters, in which traffic is highly bursty, even at microsecond granularity [15]. To cope with the limited buffer size of network switches, recent work has explored more intelligent buffer-sharing mechanisms that can dynamically allocate memory across ports and priorities to reduce packet loss during congestion [27, 2, 5, 19]. Yet, this dynamic sharing introduces its challenges. When one queue grows aggressively, it can starve or throttle others, causing unfairness, throughput degradation, and tail latency inflation. Critically, such interference can occur even across logically independent queues, whether they originate from different applications or are mapped to separate output ports [5]. Datacenter switches rely on two primary mechanisms to manage packet admission and buffer allocation: Buffer Management (BM) and Active Queue Management (AQM). BM schemes such as Complete Sharing (CS) [7], Dynamic Threshold (DT) [12], and Enhanced Dynamic Threshold (EDT) [24] aim to improve fairness in buffer utilization across output queues and reduce packet drop rates during congestion. These policies regulate buffer access to ensure that no single queue monopolizes shared memory, helping balance flow treatment under contention. In contrast, AQM algorithms–including RED [14], CoDel [22], and PIE [23], operate by monitoring queue occupancy and proactively dropping packets before
buffers overflow. This early drop behavior prevents persistent congestion and keeps queuing delays bounded, improving tail latency and responsiveness. However, these two control schemes typically operate in isolation. While BM focuses on spatial fairness across queues, AQM targets temporal responsiveness within a single queue. Lack of coordination between them can lead to undesirable side effects such as queue starvation, misallocated buffer resources, and missed opportunities for holistic congestion control. Recent proposals such as ABM [2], L2BM [19], Reverie [3], and Credence [5] have pushed the frontier of switch buffer sharing by combining BM with AQM, machine learning, or fine-grained telemetry. While effective in simulation or theoretical settings, these approaches rely on system components that are infeasible in today’s programmable switches. For example, ABM and L2BM depend on the estimation of the per-queue drain rate and floating point arithmetic to dynamically adjust thresholds, but these drain rates must be periodically calculated using continuous metrics, often every 10–30 ms, which are highly sensitive to estimation noise, especially under shared link contention or bursty flows [23, 21]. Similarly, Reverie introduces ML-based prediction logic or oracle-assisted control paths, which, while conceptually powerful, assume access to operations (e.g., push-out, priority remapping, predictive feedback) not currently supported in line-rate switch hardware. Moreover, many of these systems rely on DT to manage buffer sharing, which is known to induce buffer underutilization in real deployments due to its conservative drop behavior [12]. These limitations raise a fundamental question. Can we design a buffer sharing mechanism that is effective under dynamic traffic conditions? BS HARE answers this question by shifting complexity from the ingress path and instead uses queuing delay at the egress as a congestion signal. This design has several advantages: (i) queuing delay is directly observable in hardware and reflects real-time congestion, even under high ingress arrival rates; and (ii) delay-based signals eliminate the need for per-queue drain rate estimation or traffic prediction. In doing so, BS HARE remains simple and hardware-friendly, requiring only a single operator-configurable parameter, while matching or exceeding the performance of more complex designs. Inspired by CoDel’s delay-based philosophy, BS HARE 1 1 An early version of this idea appeared as a short position paper at a student workshop [6]. This paper presents a complete system design and comprehensive evaluation that were not part of the earlier version.
2
dynamically shares switch buffer space across queues in datacenters, operating similarly to CS [7] for buffer admission. Packets are dropped either when the buffer is full or when their queuing delay exceeds a target threshold that adapts to current buffer occupancy and congestion conditions. Our evaluation shows that BS HARE performs comparably to ABM in most scenarios and outperforms ABM by up to 45.07% in Flow Completion Time (FCT) when running modern transport protocols like PowerTCP [4]. Our contributions can be summarized as follows. • We propose BS HARE , a lightweight buffer sharing mechanism that leverages queueing delay inspired by the CoDel AQM scheme to dynamically share buffer space across queues. • We conduct extensive ns-3 simulations demonstrating that BS HARE significantly improves flow completion time (FCT) slowdown compared to static buffer management (BM) schemes. Moreover, BS HARE outperforms adaptive buffer management (ABM) approaches when combined with modern datacenter transport protocols. • We perform a detailed steady-state analysis to characterize the behavior, stability, and buffer allocation dynamics of BS HARE under varying traffic conditions. II. BACKGROUND AND M OTIVATION A. Background A shared memory switch architecture is a widely adopted design in high-performance networking devices, as illustrated in Figure 1, including switches and routers, to efficiently manage network packet flow. A central feature of this architecture is a buffer memory accessible by multiple components within the switch. Incoming packets at input ports are directed to ingress queues for temporary storage, which regulates entry into the switch fabric. Egress queues at output ports similarly store packets before transmission onto the outgoing network. BM algorithms are employed to allocate buffer space efficiently among these queues. The shared buffer memory is a critical resource, providing rapid access and storage for packets from multiple ingress and egress points simultaneously. This architecture typically incorporates a memory management unit (MMU) that arbitrates access to the shared memory, ensures fairness through BM and AQM, and optimizes throughput. Leveraging shared memory resources enables the switch to accommodate fluctuating traffic loads and maintain high throughput, a crucial aspect of modern network performance. Switches play a central role in forwarding packets across modern networks. However, packet forwarding within switches faces several challenges, including congestion, buffer contention, and inefficient queue management, all of which can significantly impact network performance. In this section, we discuss the key factors affecting switch performance, with a particular focus on congestion, starvation, and queueing delay, and analyze their impact on packet forwarding efficiency and application performance. • Congestion. Congestion in switch queues refers to a situation where the amount of incoming traffic exceeds
the capacity of the switch’s output ports or the processing capabilities of the switch itself. Effect. It results in delays in packet forwarding, increased queuing delay, packet loss, and overall degradation of network performance [25]. • Queuing Delay. When packets arrive at a network device, such as a switch or router, they are stored in queues before being processed and forwarded. Queuing delay is the time a packet spends waiting in these queues before being transmitted. As network traffic increases and queues become longer, the queuing delay also increases. Effect. This delay can impact real-time applications such as video streaming or VoIP, causing jitter and latency issues. The other term, which is often heard and related to queueing delay, is Bufferbloat, which occurs when a network device has large buffers. Having a large buffer means it can store a large number of packets, but it leads to increased latency and delays in delivering packets [25]. • Starvation. Starvation arises when specific flows or types of traffic consistently receive lower priority or insufficient network resources compared to others. Effect. This disparity in resource allocation leads to poor performance for the starved traffic. For instance, if a network prioritizes certain applications over others, lowerpriority applications may experience persistent delays or insufficient bandwidth, resulting in a state of starvation [2]. B. Motivation In this section, we present the limitations of current switch buffer sharing algorithms and then motivate the need to design our solution. Limitations of existing designs. As datacenter switch buffers continue to shrink and traffic becomes increasingly bursty and unpredictable, buffer sharing mechanisms must evolve to remain practical and deployable. Recent proposals, such as ABM [2], improve switch buffer sharing by combining spatial (BM) and temporal (AQM) control or by leveraging optimization-based approaches. While these designs demonstrate strong performance in simulation, they rely on calculating drain rates for each queue. The calculation of drain rates involves determining continuous values, which are susceptible to estimation errors, especially when multiple queues share the same link or link capacity fluctuates [23, 21]. We argue that these designs have two key limitations: (i) Unrealistic assumptions: These schemes often depend on idealized switch behavior. For example, they require finegrained per-queue metrics such as instantaneous drain rate, which are difficult or impossible to measure accurately in hardware. Some techniques assume the availability of unsupported features like push-out, floating-point operations, or solver-based coordination. (ii) Implementation complexity: Even if theoretically implementable, these designs require tightly coupled control across multiple queues and memory regions. For example, ABM depends on continuously updated per-queue drain rates and coordinates global buffer state to make drop decisions,
3
Ingress pool
Egress pool MMU
Ingress link 1
Egress link 1
Ingress link 2
Egress link 2 Switching fabric
Ingress link 3
Egress link 3
.. .
Shared buffer memory
.. .
Figure 1: The architecture of a shared-memory switch in which different colours indicate different sources and priorities of data packets with one queue per output link. logic that is difficult to parallelize and incompatible with the match-action abstraction used in programmable switches. Therefore, we design a practical buffer sharing mechanism that avoids reliance on carefully tuned parameters while achieving performance comparable to prior buffer sharing schemes. III. S YSTEM D ESIGN We aim to create a single buffer-sharing algorithm that addresses the limitations of existing methods while incorporating the following desired features: (i) Simple to configure, BS HARE uses a single operator-configurable parameter (αp ) per priority class. (ii) Driven by queuing delay, rather than raw queue lengths, enabling more accurate congestion response and fair buffer allocation. BS HARE’s core idea is that packet queuing delay, a coarse but powerful signal, can serve as a scalable proxy for drain time and congestion, without requiring heavyweight calculations or coordination. The threshold for each queue is dynamically adjusted using local delay observations and minimal shared-state arithmetic. A. The BS HARE Algorithm BS HARE draws inspiration from the simplicity and effectiveness of packet queueing delay of CoDel [22] and the spatial efficiency of CS [7]. In an output-queued shared-memory packet switching chip, BS HARE assigns a threshold θpi (t) for dequeuing a packet with priority p and port i for any particular instance of time t. Here, p represents the service priority of the packet, distinguishing traffic classes such as short flows, incast bursts, or background flows, each of which may tolerate different levels of queuing delay. This threshold enables BS HARE to adaptively regulate buffer usage based on real-time congestion signals while maintaining hardware implementability. Assume that B denotes the total available buffer space in the switch, and Q(t) represents the amount of buffer used to enqueue current packets across all ports and queues. BS HARE computes the delay threshold θpi (t) using a configurable parameter αp , the port bandwidth C, and two dynamic system metrics: (i) the number of congested queues at priority level p, denoted by cp , and (ii) the available buffer space, given by B − Q(t). The delay threshold is then defined as:
θpi (t) =
αp · (B − Q(t)) cp · C
(1)
αp is the only operator-configurable parameter. It scales the aggressiveness of delay adaptation: a larger αp increases a queue’s chance of accepting packets under congestion, similar to the role of the delay target in DT [12]. This flexibility allows tailoring queue behavior based on service-level requirements or traffic priorities. C denotes the port bandwidth. A higher bandwidth implies faster draining of queues, allowing less time to hold packets before dequeuing. As a result, the delay threshold is inversely proportional to C, ports with higher bandwidths receive proportionally smaller delay thresholds to maintain timely queue management. (B − Q(t)) captures the available buffer space at time t, representing how much of the buffer is currently unused. It reflects the system’s willingness to admit more packets: the more free space, the more permissive the threshold. Similar to DT [12], this design choice for available buffer space ensures that queues with more headroom are less aggressive in dropping packets. For example, when the buffer is half full, the threshold is twice as large as when it is nearly full, encouraging smoother utilization under varying load. cp denotes number of congested queues at priority level p. A queue is considered congested when its length nears its threshold. In BS HARE, we classify a queue as congested when its length reaches at least 90% of the overall buffer space. The threshold is inversely related to cp : as more queues become congested, the delay budget is divided among them, reducing the threshold and making each queue more conservative in accepting packets. B. Operational Overview We now present BS HARE and explain how it works. Upon arrival at the output port, BS HARE performs an admission check before enqueueing the packet. Specifically, it verifies whether accepting the packet would exceed the switch’s remaining buffer capacity B − Q(t). This check prevents buffer overflow and ensures stable queue behavior. Following a similar philosophy to CS [7], BS HARE compares the current buffer occupancy with the total available buffer of the switch. If sufficient space is available, the packet
4
is admitted to the queue while retaining its original priority. Additionally, BS HARE records the packet’s arrival timestamp during dequeuing. If the buffer is full, BS HARE immediately drops the packet to avoid further congestion, maintaining backpressure and fairness across competing queues. When packets accumulate in output (egress) queues, BS HARE applies a delay-aware dequeuing policy to determine whether a packet should be transmitted or dropped. This decision is based on two metrics: (i) the packet’s sojourn time, i.e., the time spent in the queue since arrival, and (ii) a dynamically computed target delay threshold, θpi (t), as defined in Eq. 1. If the sojourn time of a packet is below the target delay, the packet is forwarded for transmission. Otherwise, it is dropped, signaling excessive queuing and potential congestion. This delay-aware dropping mechanism ensures that stale packets do not occupy limited buffer space at the expense of newer, potentially more latency-sensitive traffic. Datacenter traffic is often bursty at sub-RTT timescales, particularly during the initial phases of a connection when congestion control has not yet reacted. To protect short flows, which are highly sensitive to even a single packet drop, BS HARE assigns higher priority (via a larger α) to packets within their first RTT. This reduces their likelihood of being dropped under transient congestion events, such as incast bursts. In contrast, lower priority traffic, such as long-lived flows, is assigned a smaller α, resulting in a smaller target delay threshold. Packets from these flows that persist too long in the buffer are more likely to be dropped, enabling TCP to respond with backoff and alleviating pressure on shared buffers. This design balances throughput efficiency with latency sensitivity across diverse flow types. Threshold update. In BS HARE, the delay threshold θpi (t), defined in Eq. 1, is updated upon every packet departure. This ensures threshold decisions remain responsive to real-time buffer dynamics while avoiding excessive update overhead. The parameters αp (priority weight), B (total buffer capacity), and C (port bandwidth) are static during runtime, offering stability and reducing tuning complexity. However, the buffer size Q(t) (used buffer space), is subject to continuous change after every packet enqueue and dequeue operation. The parameter cp , denoting the number of congested queues at priority level p, is updated less frequently, once per RTT. This amortized update strategy provides a coarse but effective congestion signal without incurring per-packet processing overhead. Safeguard. BS HARE enforces boundary conditions on buffer occupancy to maintain robust and predictable behavior under varying load. These safeguards prevent premature drops and uncontrolled buffer growth. If the queue length falls below a predefined minimum threshold (MinBytes), BS HARE suppresses packet drops—regardless of queuing delay or priority. This ensures that buffers are not underutilized, avoiding unnecessary packet loss when sufficient capacity is available. Conversely, if the buffer occupancy approaches its maximum capacity, BS HARE immediately drops incoming packets to prevent overflow. This early rejection serves as a backpressure signal, helping to mitigate congestion and preserving stability across competing queues. Together, these boundary checks
Algorithm 1: BS HARE algorithm Input: None Output: packet - Packet to transmit. /* Perform round-robin scheduling over all queues */ 1 foreach queue in Queues do 2 packet ← queue.dequeue(); 3 if packet ̸= null then 4 priority ← packet.getPriority(); 5 targetDelay ← θpi (t); /* sojournTime is the total time spent by packet in queue */ 6 sojournTime ← CurrentTime() − packet.getTimestamp(); 7 if sojournTime < targetDelay or QueueLength(priority) < MinBytes then /* Admit packet for transmission */ 8 return packet; 9 end 10 else /* Drop the packet and continue */ 11 DropAfterDequeue(packet); 12 end 13 end 14 end 15 return packet;
ensure that BS HARE remains efficient and resilient under both light and heavy traffic conditions. Relation to CoDel: While BS HARE draws conceptual inspiration from CoDel’s use of queuing delay as a congestion signal, the two serve fundamentally different roles. CoDel is an AQM algorithm designed to operate within a single queue, aiming to control latency by proactively dropping packets based on sojourn time and fixed control intervals. It uses a static target delay (typically 5ms) and adjusts its drop rate using an inverse square-root backoff mechanism over 100ms intervals. In contrast, BS HARE is a buffer sharing mechanism designed for switches with multiple output queues and shared onchip memory. It does not attempt to manage congestion within a single queue but instead determines buffer admission and packet drop decisions across queues, using a dynamic, delaybased threshold. Unlike CoDel, BS HARE requires no interval tracking or static parameters, and it is explicitly designed for efficient implementation in programmable switch data planes. BS HARE addresses several key limitations of CoDel’s packetdropping strategy: (i) Dynamic threshold adaptation: Unlike CoDel’s fixed delay target, BS HARE computes its delay threshold based on current buffer occupancy and queue contention, allowing it to adapt to varying traffic conditions in real time. (ii) Simplified control logic: CoDel requires interval track-
5
ing and square-root-based drop pacing. BS HARE avoids this complexity, using a stateless, threshold-based check on every dequeued packet. (iii) Faster congestion reaction: CoDel may react slowly to short bursts due to its interval-based design [18]. BS HARE enables immediate drop decisions based on perpacket delay, making it more responsive under bursty datacenter traffic. CoDel and BS HARE serve different purposes, AQM versus buffer management, so we will exclude CoDel in our evaluation. They cannot be interchanged in deployment, and comparing them would confuse congestion control within a queue with buffer allocation across queues. IV. S TEADY S TATE A NALYSIS In this section, we analyze the steady state of BS HARE inspired by ABM [2]. Before checking for a steady state, let us consider the model and formalize its allocation. A. Formalizing BS HARE allocation In our model, we consider a switch with a fixed number of ports and one queue per priority per port. The switch operates on a shared memory architecture with a total buffer space of B. At any given time t, the occupied buffer space is denoted as Q(t). Our analysis relies on a fluid model, assuming deterministic and continuous arrivals and departures of packet bits. For BS HARE, we use the parameter αp to allocate buffer space for each priority p. Each priority is associated with a dedicated queue at each port. Port indices are denoted by i, and p represents the priority. The set of priorities utilizing the buffer is denoted as ℘. The number of congested queues for a priority p at time t is represented by cp (t). As described in the Equation 1, the threshold time of a packet during dequeue time at the port i and belonging to priority p is determined by the alpha parameter αp , the number of congested queues cp (t), port bandwidth C, and the remaining buffer space B − Q(t). This calculation can be formally expressed as: Ωideq,p (t) =
αp .λp (t).(B − Q(t)) C
(2)
where λp (t) = is the inverse of the total number of congested queues of priority p at time t. We are introducing a new parameter delta ∆ip which is the product of αp and λp (t) having priority p at port i at instance time t. 1 cp (t)
Proof: Observing that λp (t) represents the number of congested queues for a priority p, it remains constant across all queues of the same priority; therefore, the equation is as follows: X
∆ip (t) =
i
X
αp .λp (t) =
i
X
αp .
i
cp (t) 1 = αp . = αp cp (t) cp (t) (5)
C. Analysis In the BS HARE context, a steady state refers to a situation where the load conditions and buffer occupancy remain constant. Essentially, this means that the number of packets entering the system equals the number of packets leaving it, maintaining a balance. In a steady state, each packet has a threshold time to leave the system, which represents the ideal time after which the packet will be discarded if it remains in the buffer. During a steady state, these ideal times for packets remain consistent. During the steady state of BS HARE, our primary concern is to evaluate important metrics. These include the overall buffer allocation, denoted as Q and represented by Equation 6. Additionally, we are interested in determining the remaining buffer space, denoted as B − Q, which is calculated using Equation 7. Moreover, we focus on calculating the thresholds per congested queue, represented as Ωideq,p , as specified by Equation 8, using BS HARE. P P B i p ∆ip P P (6) Q= 1 + i p ∆ip 1+
B P P
(7)
(1 +
B.∆i P Pp
(8)
B−Q= Ωideq,p =
i p ∆p
i
i
i p ∆p )
Proof 1: In the steady-state scenario, where we assume that the packet ideal time in each queue is equal to their thresholds, we can derive the overall buffer occupancy by summing the queue lengths of all congested queues of all ports and priorities. This can be represented as follows: Q X X ∆ip .(B − Q) = C C p i
(9)
Solving this equation for q leads to the final equality, which encapsulates the determination of overall buffer occupancy.
(3)
Proof 2: For remaining buffer allocation, we subtract Equation 6 from total buffer (B).
The total instantaneous sum of ∆ip (t) across all queues within a priority p ∈ ℘ across all ports is constrained by an upper bound denoted as αp .
Proof 3: To calculate thresholds per congested queue, we put the value of the remaining buffer from Equation 7 in the below threshold calculation of BS HARE and multiply by bandwidth (b) to convert to the threshold length of the queue.
∆ip (t) = αp .λp (t) B. Property of Delta
X i
∆ip (t) ≤ αp
(4)
Ωideq,p =
αp (B − Q).C cp C
(10)
6
D. Reduced dependency on scheduling algorithms The reduction of dependency on scheduling algorithms is highlighted in Equation 1. It is important to note that this equation does not rely on factors such as link rates, traffic loads, queue length, or drain rate. This reduces reliance on variables that are beyond the control or prediction of the local buffer. In a recent approach proposed by [2], however, the drain rate is incorporated into the calculation. This introduces complexities as determining drain rates involves continuous value estimation over time, which can be prone to errors, especially in scenarios where multiple queues share the same link or when link capacity fluctuates [23, 21]. E. Minimal Buffer Waste In BS HARE, we do not impose strict boundaries like CS for buffers, which allows the utilization of the entire buffer capacity while maintaining fairness among queues, as observed in previous studies [11, 29], but this differs from DT [12] and ABM [2], which allocate buffer space for transient conditions and bursty cases, resulting in potential underutilization when there are no bursty scenarios. However, CS can lead to drawbacks, particularly when long flows occupy the buffer completely and cause packet drops for shorter flows, resulting in poor performance in terms of FCT slowdown [9]. Therefore, BS HARE strikes the right balance by prioritizing short flows and bursts dynamically without sacrificing buffer underutilization, considering time as a factor. BS HARE makes sure of full buffer utilization, but in case the size of the buffer is not a multiple of packet size, there is a slight amount of buffer remaining, which is less than the packet size.
Minimum Buffer ≥
•
Isolation Drawing from ABM’s emphasis on isolation across system priorities, BS HARE similarly prioritizes equitable resource distribution through tailored mechanisms. BS HARE achieves this by (i) imposing constraints on the total buffer allocation for each priority and (ii) guaranteeing minimum buffer allocations to individual priorities. This ensures that no single priority can monopolize the buffer at the expense of others. BS HARE dynamically adjusts thresholds per queue based on congestion levels within each priority, mirroring ABM’s approach. As congestion increases within specific priority queues, BS HARE reduces thresholds for each queue within that priority, thereby promoting fair resource allocation across priorities. By applying the concept of isolation to BS HARE, it remains effective, as evidenced by the equations of 11 and 13. The equations below affirm BS HARE’s commitment to ensuring equitable resource distribution among system priorities. Minimum guarantee buffer The minimum available buffer size for any priority level p can be expressed as:
(11)
p∈℘ αp )
Proof: The threshold Ωideq,p for each queue with priority p is determined by Equation 8. Summing across all ports, we compute the total allocated buffer as follows: X
Ωideq,p =
i
(1 +
P B. i ∆ip P P i
i p∈℘ ∆p )
The last inequality holds since from Equation 4.
≥
(1 +
p∈℘ αp )
i p ∆p ≤
P P i
B.α P p
P (12) p αp
Buffer preventing monopoly The maximum available buffer size for any priority level p can be expressed as: Maximum Buffer ≤
B.αp (1 + αp )
(13)
Proof: The proof follows a similar approach to Theorem 12. To the upper bound, we leverage the P establish P property i p∈℘ ∆ip ≥ αp , which occurs when only priority p is utilizing the buffer. • Bounded drain time. BS HARE effectively manages drain time by allocating buffer space in proportion to the drain rate of individual queues. This mechanism ensures that BS HARE limits the queuing delay and the overall buffer drain time, regardless of the number of congested queues or the scheduling policy in use. The thresholds established by BS HARE provide upper bounds for the drain time τ of any queue with priority p, as given by: τ≤
F. BS HARE: Assessing Property Alignment In this section, we analyse the extent to which BS HARE adheres to the principles of ABM properties [2].
(1 +
B.α P p
B.αp C(1 + αp )
(14)
It should be noted that the upper bound of τ is only dependent on the constant parameters. Proof: Using Equation 8 from our steady-state analysis and considering that drain time is the occupied buffer divided by its bandwidth (C), we derive the drain time τ as follows for a queue at port i and of priority p B.αp . c1p B.αp P P ≤ (15) C(1 + i p ∆ip ) C(1 + αp ) P P i The last inequality holds since i p ∆p ≥ αp and 1 ≤ 1 cp τ=
V. P ERFORMANCE E VALUATION We simulate BS HARE through packet-level simulations in NS-3 and compare it against state-of-the-art buffer management schemes tailored for datacenter networks. Our simulations use a leaf-spine datacenter topology comprising eight spine switches and 256 servers connected via eight leaf switches. All links operate at 10Gbps and incur a 10us propagation delay. The network follows a 4:1 oversubscription ratio, mirroring the configuration used in prior work [2]. Switches in both the leaf and spine layers are provided with 9.6KB of
IB (AFD + Elephant trap) ABM BShare
150 100 50 0
20
40 60 Load (%)
(a) Incast Flows
80
150
DT FAB CS
100
100
IB (AFD + Elephant trap) ABM BShare
50 25 15 3
20
40 60 Load (%)
80
(b) Short Flows
DT FAB CS
80
IB (AFD + Elephant trap) ABM BShare
Avg. Throughput (%)
DT FAB CS
200
99-pct Buffer (%)
250
99-pct FCT slowdown
99-pct FCT slowdown
7
60 40 20 0
20
40 60 Load (%)
80
(c) Buffer Occupancy
100
DT FAB CS
80
IB (AFD + Elephant trap) ABM BShare
60 40 20 0
20
40 60 Load (%)
80
(d) Throughput
Figure 2: Comparison of buffer management (BM) schemes under varying network loads. BS HARE achieves performance comparable to ABM and outperforms other BM schemes in key scenarios: (a) incast flows that generate bursty traffic, (b) short flows representative of web search workloads. In addition, BS HARE achieves lower buffer occupancy (c), and sustains high throughput without degradation (d). buffer per port per Gbps, consistent with the characteristics of the Broadcom Trident II ASIC [27, 17]. We evaluate BS HARE under two representative datacenter traffic patterns. First, we use the web search workload from [8], which captures realistic flow size distributions observed in production datacenters. This workload is used to evaluate performance across a range of network loads, varying from 20% to 80%. Second, we generate incast traffic patterns, following the methodology in [2], to emulate bursty query-response behavior typical of distributed storage systems. In each incast scenario, a single query triggers simultaneous responses from multiple servers. We configure each server to issue two requests per second, and we vary the total burst size from 10% to 100% of the switch’s buffer capacity to stress the system under varying degrees of congestion. To evaluate performance under different congestion control strategies, we run experiments using two widely studied transport protocols: DCTCP [8] and PowerTCP [4]. These protocols play a crucial role in how data is transmitted and managed across the network, allowing us to evaluate their performance under different traffic conditions. Benchmarks. We compare BS HARE with five representative BM schemes drawn from both academic literature and industry practice: Active Buffer Management (ABM) [2], Dynamic Thresholds (DT) [12], Flow-Aware Buffering (FAB)[10], Complete Sharing (CS), and Cisco Intelligent Buffering (IB) [1]. Across the board, ABM allocates buffer space based on the normalized drain rate and the proportion of remaining buffer space. DT uses a simpler approach, allocating buffer strictly according to remaining capacity. FAB extends DT by incorporating flow-awareness, giving preferential treatment to short flows. CS allows any queue to grow as long as buffer space is available, emulating an aggressive complete sharing policy. IB combines DT with approximate fair dropping through a hierarchical scheme, reflecting buffer management heuristics deployed in commercial Cisco switches. To understand how BS HARE interacts with different congestion control strategies, we evaluate it alongside several transport-layer protocols: Cubic [16] (loss-based), DCTCP [8] (ECN-based), TIMELY [20] (RTT-gradient-based), PowerTCP, and θ-PowerTCP [4] (power-aware). Due to space constraints,
we present representative results that capture the key trends across these protocols. We report on three primary performance metrics: total switch buffer occupancy, throughput, and FCT slowdown. We calculate FCT slowdown as the ratio between the actual FCT and the ideal FCT measured in the absence of competing traffic, providing a normalized view of flow latency under load. Parameters configuration. We set α to 0.5 for BS HARE, ABM, DT, and FAB. For DCTCP, PowerTCP, and θPowerTCP parameters were set as per [8, 4]. We update the values of cp and the ABM drain rate once per RTT. 1) Incast Traffic FCT’s: BS HARE significantly reduces the 99th percentile FCT slowdown in incast scenarios compared to traditional buffer management schemes. Figure 2a presents results under Cubic transport with a fixed request size equal to 30% of the switch buffer, across varying network loads. At low load (20%), BS HARE performs similarly to existing schemes, including DT, FAB, CS, IB, and ABM. However, as load increases, BS HARE consistently outperforms DT, FAB, CS, and IB, and closely tracks ABM’s performance. At 40% load, BS HARE reduces FCT slowdown by an average of 88.86% compared to DT, FAB, CS, and IB. At 80% load, this gap further widens, with a 93.7% reduction in FCT slowdown, while still maintaining performance comparable to ABM. These results highlight BS HARE’s robustness under highincast pressure and demonstrate its effectiveness in preserving latency-sensitive performance, especially under aggressive buffer contention. Figure 3a shows the 99th percentile FCT slowdown for incast flows under a fixed 40% web search workload, as the incast request size varies. Even with small bursts—e.g., 12.5% of the buffer size—BS HARE delivers a meaningful reduction in FCT slowdown. On average, it improves FCT slowdown by 25.85% compared to DT, FAB, CS, and IB, and performs comparably to ABM. As the size of the request increases, the benefits of BS HARE become more apparent. For example, at a request size of 50% of the buffer size, BS HARE shows a remarkable reduction in the FCT slowdown for the incast workload, averaging at least 85.21% compared to DT, FAB, CS, and IB, and again close to ABM. The results reveal that as network load increases, BS HARE consistently outperforms traditional buffer management schemes, including DT, FAB, CS, and IB, and closely
100 50 20 10 10
25 50 75 Request Size (% of buffer size)
(a) Incast Flows
50 40
DT FAB CS
IB (AFD + Elephant trap) ABM BShare
30 20 10 1 10
25 50 75 Request Size (% of buffer size)
(b) Short Flows
100 80
DT FAB CS
IB (AFD + Elephant trap) ABM BShare
60 40 20 0 10
25 50 75 Request Size (% of buffer size)
(c) Buffer Occupancy
Avg. Throughput (%)
IB (AFD + Elephant trap) ABM BShare
99-pct Buffer (%)
DT FAB CS
150
99-pct FCT slowdown
99-pct FCT slowdown
8
100 80
DT FAB CS
IB (AFD + Elephant trap) ABM BShare
60 40 20 0 10
25 50 75 Request Size (% of buffer size)
(d) Throughput
Figure 3: Comparison of BM schemes under various bursts size. BS HARE demonstrates comparable performance to ABM and superior performance compared to other BM schemes: (a) for flows contributing to bursts (incast traffic), and (b) for short flows (websearch), across different bursts. Additionally, BS HARE utilizes less buffer (c), and maintains throughput (d). approaches the performance of ABM in all evaluated scenarios. 2) Buffer Utilization: BS HARE efficiently allocates the available buffer space of the switch for medium and long flows, even under buffer-intensive transport protocols such as Cubic. As shown in Figure 2c, it reduces average buffer occupancy by 41.5% across varying loads compared to traditional BM schemes, while maintaining throughput comparable to ABM. In addition to overall efficiency, BS HARE adapts buffer usage to accommodate bursty traffic. Figure 3c illustrates that as the incast request size increases, BS HARE strategically utilizes more buffer to absorb bursts. For large request sizes, BS HARE consumes 12.92% more buffer than DT, FAB, and IB, and 7.18% more than ABM, highlighting its ability to dynamically allocate resources when needed without compromising performance. BS HARE reduces average buffer utilization compared to traditional schemes while dynamically allocating additional buffer during burst events–demonstrating both efficiency and responsiveness. In high-burst scenarios, it surpasses ABM in buffer usage to better absorb transient congestion. 3) Throughput: We evaluate BS HARE’s throughput performance under both varying network loads and incast burst sizes. As shown in Figure 2d, BS HARE sustains throughput comparable to existing buffer management schemes across all load levels. Even under high burst pressure—illustrated in Figure 3d with large incast request sizes—BS HARE maintains stable throughput while continuing to deliver low FCTs for short flows. BS HARE preserves throughput on par with state-of-the-art schemes, while upholding delay-based thresholds and achieving low tail latency for short flows. 4) Distinct Priorities: We evaluate the performance of BS HARE across distinct priority classes using a scenario in which Cubic, DCTCP, and θ-PowerTCP flows are each assigned separate queues for both web search and incast workloads. Figure 4 shows that BS HARE significantly improves FCT slowdown compared to DT and ABM under varying load conditions. For Cubic flows, as load increases, BS HARE maintains competitive performance and begins to outperform ABM. At 50% load, it reduces FCT slowdown by 11.80% relative to ABM and by 63.72% relative to DT. In the case of θ-
PowerTCP, BS HARE consistently outperforms both baselines– achieving a 45.07% reduction in FCT slowdown at 10% load and 28.89% at 30% load compared to ABM, and over 93.44% compared to DT. For DCTCP flows, BS HARE continues to show strong performance, reducing FCT slowdown by up to 34.73% over ABM and 52.79% over DT. These improvements highlight BS HARE’s ability to prioritize flows effectively across multiple congestion control algorithms and traffic types. In distinct priority scenarios, BS HARE consistently outperforms DT and often surpasses ABM, with particularly strong gains under θ-PowerTCP across all load levels. 5) Realistic Buffer: We evaluate BS HARE under realistic buffer constraints, focusing on a baseline of 9.6KB of buffer per port per Gbps, representative of Broadcom Trident II switches [17]. To assess generality and hardware applicability, we also simulate smaller, implementable buffer sizes as seen in Tomahawk and Tofino switches [27]. We consider both web search and incast workloads at 40% network load, fixing the incast request size to 25% of the buffer size defined for Trident II. Evaluations are conducted using DCTCP and PowerTCP transport protocols. Figure 5 presents the 99th percentile FCT slowdown for incast flows under different buffer sizes. Under DCTCP (Figure 5a), BS HARE maintains stable performance across all configurations, with only a minor increase in FCT slowdown observed on Tofino. In contrast, both DT and IB degrade sharply when buffer sizes drop below 7KB per port per Gbps– showing up to a 10× increase in FCT slowdown relative to BS HARE, particularly with Tomahawk and DCTCP. Under PowerTCP (Figure 5b), DT and IB appear more resilient until buffer sizes fall to 6KB per port per Gbps. However, their limitations become evident at 5.12KB, where both schemes fail to maintain low tail latencies. BS HARE, by contrast, continues to deliver near-ABM performance across all buffer sizes tested. Across a range of realistic hardware buffer sizes, BS HARE closely tracks ABM’s performance under both DCTCP and PowerTCP. It remains robust even when buffer capacity is severely constrained, where other schemes like DT and IB suffer substantial tail latency degradation.
BShare
50 20 10 5 1
20
40 Cubic Load (%)
60
100
DT ABM
102
BShare
100
DT ABM
100
BShare
99-pct buffer (%)
DT ABM
99-pct FCT slowdown
100
99-pct FCT slowdown
99-pct FCT slowdown
9
98 12
50
10
20 10 5 1
(a) Cubic Flows
20
40 Cubic Load (%)
60
(b) DCTCP Flows
8 6 4
20
40 Cubic Load (%)
60
(c) θ-PowerTCP
DT ABM
BShare
80 60 40 20 0
20
40
60
(d) Buffer occupancy
ABM DT
103
IB (AFD + Elephant trap) BShare
102 101 10
0
2 nt 8KB de
i Tr
B 7K
B 6K
k o aw fin ah To m o T
(a) DCTCP
99-pct FCT slowdown
99-pct FCT slowdown
Figure 4: The behaviour of Cubic, DCTCP, and θ-PowerTCP when flows use separate queues, each managed by one of three BM algorithms: DT, ABM, and BS HARE. BS HARE and ABM achieve a similar FCT slowdown while BS HARE uses slightly more buffer space.
ABM DT
103
IB (AFD + Elephant trap) BShare
102 101 100 t2
en
id Tr
B
8K
B
7K
B
6K
k o aw ofin T
ah m To
(b) PowerTCP
Figure 5: Comparison of buffer management (BM) schemes under realistic buffer sizes across Trident II, Tomahawk, and Tofino switch configurations. BS HARE consistently maintains low 99th percentile FCT across buffer sizes, with the exception of a slight degradation on Tofino under DCTCP. Under PowerTCP, BS HARE achieves lower tail latency than DT and IB, and performs on par with ABM.
VI. R ELATED W ORK Switch buffer management has been widely studied in the literature via different approaches. We categorized them into hardware-based, BM, AQM, and hybrid techniques. Hardware-based. Hardware-based solutions aim to manage the buffer directly in the switch. However, they often require additional hardware support and may have limitations in handling asymmetric traffic. For instance, R-ACK [13] and ABQ [30] offer innovative approaches to control transmission rates and manage ACK packets, but their implementation complexity and dependency on hardware support pose challenges. BM-based. BM algorithms dynamically allocate buffer space among different queues or flows within network devices to prevent congestion and ensure fair distribution. Example algorithms are PO [26, 28], DT [12], and EDT [24] that aim to excel in buffer space efficiency and burst absorption. However, implementing some BM algorithms, like PO, can be complex and may require additional hardware support. AQM-based. AQM algorithms focus on regulating queue lengths by selectively admitting or dropping incoming packets based on predefined thresholds. The AQM schemes like PIE [23], CoDel [22], and RED [14] effectively handle bursty traffic and aim to maintain low queue lengths to prevent congestion. However, they may face challenges in switch fabric implementation due to their reliance on calculating average queue length and drain rate.
Hybrid-based. ABM [2] combines the strengths of BM and AQM algorithms to achieve robust isolation properties and stable buffer drain times. ABM offers efficient burst absorption and high predictability while maintaining high throughput. Although ABM exhibits promising performance improvements over existing techniques, calculating the drain rate for each queue remains challenging, particularly in fluctuating link capacities. Unlike ABM, BS HARE does not need the drain rate calculation. L2BM [19] has a similar logic to ABM. Credence [5] proposes a drop-tail buffer sharing algorithm with machine learning-based prediction to improve performance. VII. C ONCLUSION In this paper, we presented BS HARE, a practical and delayaware buffer sharing mechanism for datacenter switches, inspired by CoDel’s queuing delay signals. BS HARE is designed to address the challenges of dynamic buffer allocation in shared-memory switches using a lightweight, single-parameter control policy. Unlike prior approaches that rely on complex rate estimation or multi-stage coordination, BS HARE requires no predictive modeling. Through extensive simulation, we show that BS HARE consistently reduces flow completion times across a range of traffic loads, buffer sizes, and transport protocols—including under burst-heavy workloads and with advanced congestion control like PowerTCP. R EFERENCES [1] https://www.ciscolive.com/c/dam/r/ciscolive/apjc/docs/ 2018/pdf/BRKDCT-3640.pdf, 2018. [2] Vamsi Addanki, Maria Apostolaki, Manya Ghobadi, Stefan Schmid, and Laurent Vanbever. Abm: active buffer management in datacenters. In Proceedings of the ACM SIGCOMM 2022 Conference, SIGCOMM ’22, page 36–52, 2022. [3] Vamsi Addanki, Wei Bai, Stefan Schmid, and Maria Apostolaki. Reverie: Low pass Filter-Based switch buffer sharing for datacenters with RDMA and TCP traffic. In 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24), pages 651–668, Santa Clara, CA, April 2024. USENIX Association. [4] Vamsi Addanki, Oliver Michel, and Stefan Schmid. {PowerTCP}: Pushing the performance limits of datacenter networks. In 19th USENIX symposium on networked
10
systems design and implementation (NSDI 22), pages 51– 70, 2022. [5] Vamsi Addanki, Maciej Pacut, and Stefan Schmid. Credence: Augmenting datacenter switch buffer sharing with ml predictions. In 21st USENIX symposium on networked systems design and implementation (NSDI 24), 2024. [6] Krishna Agrawal, Vamsi Addanki, and Habib Mostafaei. Dequeue rate-agnostic switch buffer sharing through packet queueing delay. In Proceedings of the CoNEXT on Student Workshop 2024, CoNEXT-SW ’24, page 1–2, 2024. [7] William Aiello, Alex Kesselman, and Yishay Mansour. Competitive buffer management for shared-memory switches. ACM Trans. Algorithms, 5(1), dec 2008. [8] Mohammad Alizadeh, Albert Greenberg, David A. Maltz, Jitendra Padhye, Parveen Patel, Balaji Prabhakar, Sudipta Sengupta, and Murari Sridharan. Data center tcp (dctcp). In Proceedings of the ACM SIGCOMM 2010 Conference, SIGCOMM ’10, page 63–74, 2010. [9] Maria Apostolaki, Vivek Addanki, Mohammad Ghobadi, and Laurent Vanbever. FB: a flexible buffer management scheme for data center switches. 2021. [10] Maria Apostolaki, Laurent Vanbever, and Manya Ghobadi. Fab: Toward flowaware buffer sharing on programmable switches. In ACM Workshop on Buffer, 2019. [11] J.W. Causey and H.S. Kim. Comparison of buffer allocation schemes in atm switches: complete sharing, partial sharing, and dedicated allocation. In Proceedings of ICC/SUPERCOMM’94 - 1994 International Conference on Communications, pages 1164–1168 vol.2, 1994. [12] A.K. Choudhury and E.L. Hahne. Dynamic queue length thresholds for shared-memory packet switches. IEEE/ACM Transactions on Networking, 6(2):130–140, 1998. [13] Xinle Du, Ke Xu, Lei Xu, Kai Zheng, Meng Shen, Bo Wu, and Tong Li. R-aqm: Reverse ack active queue management in multitenant data centers. IEEE/ACM Transactions on Networking, 31(2):526–541, 2023. [14] S. Floyd and V. Jacobson. Random early detection gateways for congestion avoidance. IEEE/ACM Transactions on Networking, 1(4):397–413, 1993. [15] Ehab Ghabashneh, Yimeng Zhao, Cristian Lumezanu, Neil Spring, Srikanth Sundaresan, and Sanjay Rao. A microscopic view of bursts, buffer contention, and loss in data centers. In Proceedings of the 22nd ACM Internet Measurement Conference, IMC ’22, page 567–580, 2022. [16] Sangtae Ha, Injong Rhee, and Lisong Xu. Cubic: a new tcp-friendly high-speed tcp variant. SIGOPS Oper. Syst. Rev., 42(5):64–74, jul 2008. [17] Broadcom Inc. Trident2 / bcm56850 series. https: //www.broadcom.com/products/ethernet-connectivity/ switching/strataxgs/bcm56850-series. Accessed: 13 March 2024.
[18] Ilpo Järvinen and Markku Kojo. Evaluating codel, pie, and hred aqm techniques with load transients. In 39th Annual IEEE Conference on Local Computer Networks, pages 159–167, 2014. [19] Yi Liu, Jiangping Han, Kaiping Xue, Ruidong Li, and Jian Li. L2bm: Switch buffer management for hybrid traffic in data center networks. In 2023 IEEE 43rd International Conference on Distributed Computing Systems (ICDCS), pages 1–11, 2023. [20] Radhika Mittal, Vinh The Lam, Nandita Dukkipati, Emily Blem, Hassan Wassel, Monia Ghobadi, Amin Vahdat, Yaogong Wang, David Wetherall, and David Zats. Timely: Rtt-based congestion control for the datacenter. In Proceedings of the 2015 ACM Conference on Special Interest Group on Data Communication, SIGCOMM ’15, page 537–550, 2015. [21] Habib Mostafaei and Georgios Smaragdakis. Per priority data rate measurement in data plane. In Proceedings of the 6th on European P4 Workshop, EuroP4 ’23, page 9–15, 2023. [22] Kathleen Nichols and Van Jacobson. Controlling queue delay: A modern aqm is just one piece of the solution to bufferbloat. Queue, 10(5):20–34, may 2012. [23] Rong Pan, Preethi Natarajan, Chiara Piglione, Mythili Suryanarayana Prabhu, Vijay Subramanian, Fred Baker, and Bill VerSteeg. Pie: A lightweight control scheme to address the bufferbloat problem. In 2013 IEEE 14th International Conference on High Performance Switching and Routing (HPSR), pages 148–155, 2013. [24] Danfeng Shan, Wanchun Jiang, and Fengyuan Ren. Analyzing and enhancing dynamic threshold policy of data center switches. IEEE Transactions on Parallel and Distributed Systems, 28(9):2454–2470, 2017. [25] Andrew S. Tanenbaum. COMPUTER NETWORKS. Pearson, 2010. [26] A. K. Thareja and A. K. Agarwala. On the design of optimal policy for sharing finite buffers. IEEE Trans. Commun., June 1984. [27] Baia Wei, Shuihai Hu, Kai Chen, Kun Tan, and Yongqiang Xiong. One more config is enough: Saving (dc)tcp for high-speed extremely shallow-buffered datacenters. IEEE/ACM Transactions on Networking, 29(2):489–502, 2021. [28] S. X. Wei, E. J. Coyle, and M. T. Hsiao. An optimal buffer management policy for high-performance packet switching. IEEE GLOBECOM’91, Dec. 1991. [29] Guo-Liang Wu and J.W. Mark. A buffer allocation scheme for atm networks: complete sharing based on virtual partition. IEEE/ACM Transactions on Networking, 3(6):660–670, 1995. [30] Lei Xu, Ke Xu, Tong Li, Kai Zheng, Meng Shen, Xiaojiang Du, and Xinle Du. Abq: Active buffer queueing in datacenters. IEEE Network, 34(2):232–237, 2020.