Damon Wischik

Improving datacenter performance and robustness with multipath TCP

The latest large-scale data centers offer higher aggregate bandwidth and robustness by creating multiple paths in the core of the net- work. To utilize this bandwidth requires different flows take different paths, which poses a challenge. In short, a single-path transport seems ill-suited to such networks. We propose using Multipath TCP as a replacement for TCP in such data centers, as it can effectively and seamlessly use available bandwidth, giving improved throughput and better fairness on many topologies. We investigate what causes these benefits, teasing apart the contribution of each of the mechanisms used by MPTCP. Using MPTCP lets us rethink data center networks, with a different mindset as to the relationship between transport protocols, rout- ing and topology. MPTCP enables topologies that single path TCP cannot utilize. As a proof-of-concept, we present a dual-homed variant of the FatTree topology. With MPTCP, this outperforms FatTree for a wide range of workloads, but costs the same. In existing data centers, MPTCP is readily deployable leveraging widely deployed technologies such as ECMP. We have run MPTCP on Amazon EC2 and found that it outperforms TCP by a factor of three when there is path diversity. But the biggest benefits will come when data centers are designed for multipath transports.

The resource pooling principle

Since the ARPAnet, network designers have built localized mechanisms for statistical multiplexing, load balancing, and failure resilience, often without understanding the broader implications. These mechanisms are all types of resource pooling whichmeans making a collection of resources behave like a single pooled resource. We believe that the natural evolution of the Internet is that it should achieve resource pooling by harnessing the responsiveness of multipath-capable end systems. We argue that this approach will solve the problems and limitations of the current piecemeal approaches.

Part I: buffer sizes for core routers

In this article we describe recent work on buffer sizing for core Internet routers. This work suggests that the widely-used rule of thumb leads to buffers which are much larger than they need to be. For example, the buffer in a backbone router could be reduced from 1,000,000 packets to 10,000 without loss in performance. It could be reduced even further, perhaps to 10-20 packets, at the cost of a small amount of bandwidth utilization. This tradeoff is worth considering, for example for a possible future all-optical router.

Part II: control theory for buffer sizing

This article describes how control theory has been used to address the question of how to size the buffers in core Internet routers. Control theory aims to predict whether the is stable, i.e. whether TCP flows are desynchronized. If flows are desynchronized then small buffers are sufficient [14 ]; the theory here shows that small buffers actually promote desynchronization--a virtuous circle.

Buffer sizes for large multiplexers: TCP queueing theory and instability analysis

In large multiplexers with many TCP flows, the aggregate traffic flow behaves predictably; this is a basis for the fluid model of Misra, Gong and Towsley V. Misra et al., (2000) and for a growing literature on fluid models of congestion control. In this paper we argue that different fluid models arise from different buffer-sizing regimes. We consider the large buffer regime (buffer size is bandwidth-delay product), an intermediate regime (divide the large buffer size by the square root of the number of flows), and the small buffer regime (buffer size does not depend on number of flows). Our arguments use various techniques from queueing theory. We study the behaviour of these fluid models (on a single bottleneck Kink, for a collection of identical long-lived flows). For what parameter regimes is the fluid model stable, and when it is unstable what is the size of oscillations and the impact on goodput? Our analysis uses an extension of the Poincare-Linstedt method to delay-differential equations. We find that large buffers with drop-tail have much the same performance as intermediate buffers with either drop-tail or AQM; that large buffers with RED are better at least for window sizes less than 20 packets; and that small buffers with either drop-tail or AQM are best over a wide range of window sizes, though the buffer size must be chosen carefully. This suggests that buffer sizes should be much much smaller than is currently recommended.

O3-04-07: Tau aggregation inhibitor (TAI) therapy with rember™ arrests disease progression in mild and moderate Alzheimer's disease over 50 weeks

Background: Neurofibrillary degeneration is correlated with clinical dementia. Methylthioninium chloride (MTC, rember) dissolves Tau polymers (Paired Helical Filaments) isolated from AD brain, and prevents Tau aggregation in cell models at the nanomolar range (0.15 0.58 M). MTC has efficacy in Tau transgenic animal models, reversing cognitive and other behavioural defects, and reversing Tau pathology in the brain. Methods: An exploratory, dose-range finding, parallel design, double-blind, randomised, placebo-controlled trial of rember monotherapy was conducted in 332 subjects meeting DSM-IV and NINCDS-ADRDA for probable AD in UK and Singapore. The primary objective was to investigate the effects of oral MTC at 30, 60 and 100 mg three times per day, compared with placebo, on cognitive function (ADAS-cog) in patients with mild or moderate AD stratified by CDR. The 100mg dose was found to have a formulation defect limiting release of the therapeutic form of MTC. Secondary outcomes included MMSE, CDRsb, CGIC, ADFACS and NPI. Nested studies of SPECTand PET-scan outcomes at 6 months are reported separately. Results: In the prespecified analysis at 24 weeks, rember produced a significant improvement relative to placebo of -5.4 ADAS-cog units in CDR-moderate subjects at the 60mg dose. There was no placebo decline in CDR-mild AD over the first 24 weeks preventing initial efficacy analysis, although efficacy was demonstrated in mild AD by the SPECTscan outcomes. rember TM stabilised the progression of AD over 50 weeks in both mild and moderate AD. The overall effect size was -6.8 ADAS-cog units vs. decline if 7.8 units in the placebo/comparator arm, with significant efficacy demonstrated separately in mild and moderate subgroups. rember efficacy was confirmed on all secondary outcomes. The rember efficacy analysis met new EMEA guidelines for disease-modifying therapy by slope analysis and two-time-point analysis, supported by SPECT and PET evidence of efficacy in brain regions strongly affected by Tau pathology. Conclusions: This represents the first evidence that TAI monotherapy with rember is a viable disease-modifying treatment for mild and moderate AD which may also have preventative application at preclinical Braak stages of AD.

Optimal Scheduling Algorithms for Input-Queued Switches

The input-queued switch architecture is widely used in Internet routers, due to its ability to run at very high line speeds. A central problem in designing an input-queued switch is choosing the scheduling algorithm, i.e. deciding which packets to transfer from ingress ports to egress ports in a given timeslot. Important metrics for evaluating a scheduling algorithm are its throughput and average delay. The well-studied ‘Maximum-Weight’ algorithm has been proved to have maximal throughput [1]; later work [2]–[4] found a wider class of algorithms which also have maximal throughput. The delay performance of these algorithms is less well understood. In this paper, we present a new technique for analysing scheduling algorithms which can explain their delay performance. In particular, we are able to explain the empirical observations in [2] about the average delay in a parameterized class of algorithms akin to Maximum-Weight. We also propose an optimal scheduling algorithm. Our technique is based on critically-balanced fluid model equations.

Admission control for booking ahead shared resources

Calls that make large, persistent demands for network resources will be denied consistent service, unless the network employs adequate control mechanisms. Calls of this type include video conferences. Although overprovisioning network capacity would increase the likelihood of accepting these calls, it is a very expensive option to apply uniformly in a large network, especially as the calls require high bandwidth and low blocking probabilities. Such large calls typically require coordination of geographically distributed facilities and people at the end systems. So it is natural to book the network requirements ahead of their actual use. We present a new, effective admission control algorithm for booking ahead network services. The admission control is based on a novel application of effective bandwidth theory to the time domain. Systematic and comprehensive simulation experiments provide an understanding of how booking ahead affects call blocking and network utilization, considering call duration, number of links, bandwidth, routing, and the mix of book ahead versus immediate arrival traffic. Allowing some calls to book ahead radically reduces their chance of service denial, while allowing flexible and efficient sharing of network resources with normal calls that do not book ahead.

Data center networking with multipath TCP

Recently new data center topologies have been proposed that offer higher aggregate bandwidth and location independence by creating multiple paths in the core of the network. To effectively use this bandwidth requires ensuring different flows take different paths, which poses a challenge. Plainly put, there is a mismatch between single-path transport and the multitude of available network paths. We propose a natural evolution of data center transport from TCP to multipath TCP. We show that multipath TCP can effectively and seamlessly use available bandwidth, providing improved throughput and better fairness in these new topologies when compared to single path TCP and randomized flow-level load balancing. We also show that multipath TCP outperforms laggy centralized flow scheduling without needing centralized control or additional infrastructure.

The output of a switch, or, effective bandwidths for networks

Consider a switch which queues traffic from many independent input flows. We show that in the large deviations limiting regime in which the number of inputs increases and the service rate and buffer size are increased in proportion, the statistical characteristics of a flow are essentially unchanged by passage through the switch. This significantly simplifies the analysis of networks of switches. It means that each traffic flow in a network can be assigned an effective bandwidth, independent of the other flows, and the behaviour of any switch in the network depends only on the effective bandwidths of the flows using it.