A switch-based approach to throughput collapse and starvation in data centers

Shpiner, Alexander; Keslassy, Isaac; Bracha, Gabi; Dagan, Eyal; Iny, O.; Soha, Eyal

doi:10.1016/j.comnet.2012.06.001

Cited by 11 publications

(8 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…On the contrary, the average loads of DCNs are not very high and most of the links are not always utilized at their full ca-timeout in TCP [6,7,11,13,[18][19][20][21][22][23][24]. The authors in [6] and [18] suggest modifying the IEEE 802.1Qau switch to regulate the flow rate so that the severe drop that causes RTO (Retransmission Timeout) is less likely to happen.…”

Section: Mptcp Incast Collapsementioning

confidence: 99%

“…The authors in [6] and [18] suggest modifying the IEEE 802.1Qau switch to regulate the flow rate so that the severe drop that causes RTO (Retransmission Timeout) is less likely to happen. Shpiner et al [19] proposed using two queues to prioritize flows with a lower packet arrival rate so that it could reduce throughput disparity and make TCP less likely to resort to RTO for recovery. Alizadeh et al [20] proposed DCTCP to allow senders to start reacting as soon as the queue length on an interface exceeds a thresh value so that more buffer space is available to absorb transient micro-bursts, greatly mitigating costly packet losses that can lead to timeouts.…”

Section: Mptcp Incast Collapsementioning

confidence: 99%

See 1 more Smart Citation

MPTCP incast in data center networks

Li¹,

Lukyanenko

Tarkoma

et al. 2014

China Commun.

View full text Add to dashboard Cite

25 Recently, new data center topologies, such as the dual-homed topology, have been proposed to offer higher aggregate bandwidth by taking advantage of multiple paths. The dual-homed topology is possible because many of the current servers are shipped with dual Gigabit Ethernet onboard. Therefore, multi-homing becomes a natural solution to add more capacity to the access links between servers and ToRs (Top-of-rack switches). However, the benefit obtained from multiple paths is limited in using additional Ethernet links if a single-path TCP is used on the servers as one TCP flow cannot utilize two paths at the same time. In order to effectively and seamlessly take the advantage of multiple paths to provide improved throughput and better fairness, MPTCP (Multipath TCP) [2,3] is proposed as a replacement for TCP in a multihomed variant of the FatTree topology [4,5].One communication pattern, termed incast by the research community, poses a critical problem in single-homed DCNs due to its catastrophic goodput collapse [6][7][8][9][10][11][12][13]. In the incast communication pattern, a receiver issues a request to multiple servers (many of them Abstract: In recent years, dual-homed topologies have appeared in data centers in order to offer higher aggregate bandwidth by using multiple paths simultaneously. Multipath TCP (MPTCP) has been proposed as a replacement for TCP in those topologies as it can efficiently offer improved throughput and better fairness. However, we have found that MPTCP has a problem in terms of incast collapse where the receiver suffers a drastic goodput drop when it simultaneously requests data over multiple servers.In this paper, we investigate why the goodput collapses even if MPTCP is able to actively relieve hot spots. In order to address the problem, we propose an equally-weighted congestion control algorithm for MPTCP, namely EW-MPTCP, without need for centralized control, additional infrastructure and a hardware upgrade. In our scheme, in addition to the coupled congestion control performed on each subflow of an MPTCP connection, we allow each subflow to perform an additional congestion control operation by weighting the congestion window in reverse proportion to the number of servers. The goal is to mitigate incast collapse by allowing multiple MPTCP subflows to compete fairly with a single-TCP flow at the shared bottleneck. The simulation results show that our solution mitigates the incast problem and noticeably improves goodput in data centers. I n t h i s pa p e r, w e present the design and evaluation of EW-MPTCP to improve MPTCP performance for the many-to-one communication pattern in DCNs. Extensive simulations show that EW-MPTCP could efficiently throttle the incast congestion in various traffic requirements.

show abstract

Section: Mptcp Incast Collapsementioning

confidence: 99%

Section: Mptcp Incast Collapsementioning

confidence: 99%

MPTCP incast in data center networks

Li¹,

Lukyanenko

Tarkoma

et al. 2014

China Commun.

View full text Add to dashboard Cite

show abstract

“…Both [30] and [31] suggest to modify a IEEE 802.1Qau [32] switch to regulate flow rates, so that severe drop that causes incast throughput collapse is less likely to happen. In [33], two queues are used to prioritize flows with lower packet arrival rate, so that it can reduce throughput disparity and make TCP less likely to resort to RTO for recovery.…”

Section: Previous Workmentioning

confidence: 99%

Preventing TCP incast throughput collapse at the initiation, continuation, and termination

Tam¹,

Xi²,

Xu³

et al. 2012

2012 IEEE 20th International Workshop on Quality of Service

View full text Add to dashboard Cite

Incast applications have grown in popularity with the advancement of data center technology. It is found that the TCP incast may suffer from the throughput collapse problem, as a consequence of TCP retransmission timeouts when the bottleneck buffer is overwhelmed and causes the packet losses. This is critical to the Quality of Service of cloud computing applications. While some previous literature has proposed solutions, we still see the problem not completely solved. In this paper, we investigate the three root causes for the poor performance of TCP incast flows and propose three solutions, one for each at the beginning, the middle and the end of a TCP connection. The three solutions are: admission control to TCP flows so that the flow population would not exceed the network's capacity; retransmission based on timestamp to detect loss of retransmitted packets; and reiterated FIN packets to keep the TCP connection active until the the termination of a session is acknowledged. The orchestration of these solutions prevents the throughput collapse. The main idea of these solutions is to ensure all the on-going TCP incast flows can maintain the self-clocking, thus eliminates the need to resort to retransmission timeout for recovery. We evaluate these solutions and find them work well in preventing the retransmission timeout of TCP incast flows, hence also preventing the throughput collapse. * While we focused on TCP NewReno (RFC2582), the discussion in this paper is valid for all other modern TCP variants because those variants differ only on the congestion control mechanism. Their flow control, i.e. loss detection and recovery procedure, remains the same.

show abstract

“…In [15], the authors explored how Quantized Congestion Notification (QCN) [16], an Ethernet layer congestion control mechanism, operated under TCP-incast scenario in data centers, and proposed some modifications to improve its efficiency. Shpiner and Keslassy [7] proposed a new architecture called Hashed Credits Fair (HCF) to avoid TCP throughput collapse in data centers. The key idea of the scheme is to serve the incoming traffic through two queues (high priority and low priority queues) instead of one traditional DropTail one.…”

Section: Influence Of the Timer Granularitymentioning

confidence: 99%

“…These approaches require changes at the transport level, or within the operating system (OS) kernel. The other potential solution is to rely on switch-based mechanisms within the network [7].…”

Section: Introductionmentioning

confidence: 99%

Solving the TCP-Incast Problem with Application-Level Scheduling

Podlesny

Williamson

2012

2012 IEEE 20th International Symposium on Modeling, Analysis and Simulation of Computer and Telecommunication Systems

View full text Add to dashboard Cite

Abstract-Data center networks are characterized by high link speeds, low propagation delays, small switch buffers, and temporally clustered arrivals of many concurrent TCP flows fulfilling data transfer requests. However, the combination of these features can lead to transient buffer overflow and bursty packet losses, which in turn lead to TCP retransmission timeouts that degrade the performance of short-lived flows. This so-called TCP-incast problem can cause TCP throughput collapse. In this paper, we explore an application-level approach for solving this problem. The key idea of our solution is to coordinate the scheduling of short-lived TCP flows so that no data loss happens. We develop a mathematical model of lossless data transmission, and estimate the maximum goodput achievable in data center networks. The results indicate non-monotonic goodput that is highly sensitive to specific parameter configurations in the data center network. We validate our model using ns-2 [1] network simulations, which show good correspondence with the theoretical results.

show abstract

A switch-based approach to throughput collapse and starvation in data centers

Cited by 11 publications

References 22 publications

MPTCP incast in data center networks

MPTCP incast in data center networks

Preventing TCP incast throughput collapse at the initiation, continuation, and termination

Solving the TCP-Incast Problem with Application-Level Scheduling

Contact Info

Product

Resources

About