Sustainable and cost-effective long-term storage remains an unsolved problem. The most widely used storage technologies today are magnetic (hard disk drives and tape). They use media that degrades over time and has a limited lifetime, which leads to inefficient, wasteful, and costly solutions for long-lived data. This article presents Silica: the first cloud storage system for archival data underpinned by quartz glass, an extremely resilient media that allows data to be left in situ indefinitely. The hardware and software of Silica have been co-designed and co-optimized from the media up to the service level with sustainability as a primary objective. The design follows a cloud-first, data-driven methodology underpinned by principles derived from analyzing the archival workload of a large public cloud service. Silica can support a wide range of archival storage workloads and ushers in a new era of sustainable, cost-effective storage.

Sustainable and cost-effective long-term storage remains an unsolved problem. The most widely used storage technologies today are magnetic (hard disk drives and tape). They use media that degrades over time and has a limited lifetime, which leads to inefficient, wasteful, and costly solutions for long-lived data. This paper presents Silica: the first cloud storage system for archival data underpinned by quartz glass, an extremely resilient media that allows data to be left in situ indefinitely. The hardware and software of Silica have been co-designed and co-optimized from the media up to the service level with sustainability as a primary objective. The design follows a cloud-first, data-driven methodology underpinned by principles derived from analyzing the archival workload of a large public cloud service. Silica can support a wide range of archival storage workloads and ushers in a new era of sustainable, cost-effective storage.

The difficulty of building large data centers in dense metro areas is pushing big cloud providers towards a different approach to scaling: multiple smaller data centers within tens of kilometers of each other, comprising a "region". We show that networking this small number of nearby sites with each other is a surprisingly challenging and multi-faceted problem. We draw out the operational goals and constraints of such networks, and highlight the design trade-offs involved using data from Microsoft Azure's regions.

Internet communication today typically involves intermediary middleboxes like caches, compression proxies, or virus scanners. Unfortunately, as encryption becomes more widespread, these middleboxes become blind and we lose their security, functionality, and performance benefits. Despite initial efforts in both industry and academia, we remain unsure how to integrate middleboxes into secure sessions---it is not even clear how to define "secure" in this multi-entity context.

The use of public cloud infrastructure for storing and processing large datasets raises new security concerns. Current solutions propose encrypting all data, and accessing it in plaintext only within secure hardware. Nonetheless, the distributed processing of large amounts of data still involves intensive encrypted communications between different processing and network storage units, and those communications patterns may leak sensitive information. We consider secure implementation of MapReduce jobs, and analyze their intermediate traffic between mappers and reducers. Using datasets that include personal and geographical data, we show how an adversary that observes the runs of typical jobs can infer precise information about their input. We give a new definition of data privacy for MapReduce, and describe two provably-secure, practical solutions. We implement our solutions on top of VC3, a secure implementation of Hadoop, and evaluate their performance.

Many network functions executed in modern datacenters, e.g., load balancing, application-level QoS, and congestion control, exhibit three common properties at the data-plane: they need to access and modify state, to perform computations, and to access application semantics -- this is critical since many network functions are best expressed in terms of application-level messages. In this paper, we argue that the end hosts are a natural enforcement point for these functions and we present Eden, an architecture for implementing network functions at datacenter end hosts with minimal network support. Eden comprises three components, a centralized controller, an enclave at each end host, and Eden-compliant applications called stages. To implement network functions, the controller configures stages to classify their data into messages and the enclaves to apply action functions based on a packet's class. Our Eden prototype includes enclaves implemented both in the OS kernel and on programmable NICs. Through case studies, we show how application-level classification and the ability to run actual programs on the data-path allows Eden to efficiently support a broad range of network functions at the network's edge.

We present VC3, the first system that allows users to run distributed MapReduce computations in the cloud while keeping their code and data secret, and ensuring the correctness and completeness of their results. VC3 runs on unmodified Hadoop, but crucially keeps Hadoop, the operating system and the hyper visor out of the TCB, thus, confidentiality and integrity are preserved even if these large components are compromised. VC3 relies on SGX processors to isolate memory regions on individual computers, and to deploy new protocols that secure distributed MapReduce computations. VC3 optionally enforces region self-integrity invariants for all MapReduce code running within isolated regions, to prevent attacks due to unsafe memory reads and writes. Experimental results on common benchmarks show that VC3 performs well compared with unprotected Hadoop: VC3's average runtime overhead is negligible for its base security guarantees, 4.5% with write integrity and 8% with read/write integrity.

Motivated by streaming and data analytics scenarios where many queries operate on the same data and perform similar computations, we propose program consolidation for merging multiple user-defined functions (UDFs) that operate on the same input. Program consolidation exploits common computations between UDFs to generate an equivalent optimized function whose execution cost is often much smaller (and never greater) than the sum of the costs of executing each function individually. We present a sound consolidation calculus and an effective algorithm for consolidating multiple UDFs. Our approach is purely static and uses symbolic SMT-based techniques to identify shared or redundant computations. We have implemented the proposed technique on top of the Naiad data processing system. Our experiments show that our algorithm dramatically improves overall job completion time when executing user-defined filters that operate on the same data and perform similar computations.

Balanced graph partitioning in the streaming setting is a key problem to enable scalable and efficient computations on massive graph data such as web graphs, knowledge graphs, and graphs arising in the context of online social networks. Two families of heuristics for graph partitioning in the streaming setting are in wide use: place the newly arrived vertex in the cluster with the largest number of neighbors or in the cluster with the least number of non-neighbors.

Cellular data networks are proliferating to address the need for ubiquitous connectivity. To cope with the increasing number of subscribers and with the spatiotemporal variations of the wireless signals, current cellular networks use opportunistic schedulers, such as the Proportional Fairness scheduler (PF), to maximize network throughput while maintaining fairness among users. Such scheduling decisions are based on channel quality metrics and Automatic Repeat reQuest (ARQ) feedback reports provided by the User's Equipment (UE). Implicit in current networks is the a priori trust on every UE's feedback. Malicious UEs can, thus, exploit this trust to disrupt service by intelligently faking their reports. This work proposes a trustworthy version of the PF scheduler (called TPF) to mitigate the effects of such Denial-of-Service (DoS) attacks. In brief, based on the channel quality reported by the UE, we assign a probability to possible ARQ feedbacks. We then use the probability associated with the actual ARQ report to assess the UE's reporting trustworthiness. We adapt the scheduling mechanism to give higher priority to more trusted users. Our evaluations show that TPF 1) does not induce any performance degradation under benign settings, and 2) it completely mitigates the effects of the activity of malicious UEs. In particular, while colluding attackers can obtain up to 77 percent of the time slots with the most sophisticated attack, TPF is able to contain this percentage to as low as 6 percent.

In the last decade we have seen a huge deployment of cheap clusters to run data analytics workloads. The conventional wisdom in industry and academia is that scaling out using a cluster of commodity machines is better for these workloads than scaling up by adding more resources to a single server. Popular analytics infrastructures such as Hadoop are aimed at such a cluster scale-out environment.

Cellular data networks are proliferating to address the need for ubiquitous connectivity. To cope with the increasing number of subscribers and with the spatio-temporal variations of the wireless signals, current cellular networks use opportunistic schedulers, such as the Proportional Fairness scheduler (PF), to maximize network throughput while maintaining fairness among users. Such scheduling decisions are based on channel quality metrics and Automatic Repeat reQuest (ARQ) feedback reports provided by the User's Equipment (UE). Implicit in current networks is the a priori trust on every UE's feedback. Malicious UEs can thus exploit this trust to disrupt service by intelligently faking their reports. This work proposes a trustworthy version of the PF scheduler (called TPF) to mitigate the effects of such Denial-of-Service (DoS) attacks. In brief, based on the channel quality reported by the UE, we assign a probability to possible ARQ feedbacks. We then use the probability associated with the actual ARQ report to assess the UE's trustworthiness. We adapt the scheduling mechanism to give higher priority to more trusted users. Our evaluations show that TPF (i) does not induce any performance degradation under benign settings, and (ii) it completely mitigates the effects of the activity of malicious UEs.

Dense, unmanaged IEEE 802.11 deployments tempt saboteurs into launching jamming attacks by injecting malicious interference. Nowadays, jammers can be portable devices that transmit intermittently at low power in order to conserve energy. In this paper, we first conduct extensive experiments on an indoor 802.11 network to assess the ability of two physical-layer functions, rate adaptation and power control, in mitigating jamming. In the presence of a jammer, we find that: 1) the use of popular rate adaptation algorithms can significantly degrade network performance; and 2) appropriate tuning of the carrier sensing threshold allows a transmitter to send packets even when being jammed and enables a receiver to capture the desired signal. Based on our findings, we build ARES, an Anti-jamming REinforcement System, which tunes the parameters of rate adaptation and power control to improve the performance in the presence of jammers. ARES ensures that operations under benign conditions are unaffected. To demonstrate the effectiveness and generality of ARES, we evaluate it in three wireless test-beds: 1) an 802.11n WLAN with MIMO nodes; 2) an 802.11a/g mesh network with mobile jammers; and 3) an 802.11a WLAN with TCP traffic. We observe that ARES improves the network throughput across all test-beds by up to 150%.

As more technologies enter the home, householders are burdened with the task of digital housekeeping-managing and sharing digital resources like bandwidth. In response to this, we created and evaluated a domestic tool for bandwidth management called Home Watcher. Our field trial showed that when resource contention amongst different household members is made visible, people's understanding of bandwidth changes and household politics are revealed. In this paper, we describe the consequences of showing real time resource usage in a home, and how this varies depending on the social make up of the household.

Hermes is an optimization engine for large-scale enterprise e-mail services. Such services could be hosted by a virtualized e-mail service provider, or by dedicated enterprise data centers. In both cases we observe that the pattern of e-mails between employees of an enterprise forms an implicit social graph. Hermes tracks this implicit social graph, periodically identifies clusters of strongly connected users within the graph, and co-locates such users on the same server. Co-locating the users reduces storage requirements: senders and recipients who reside on the same server can share a single copy of an e-mail. Co-location also reduces inter-server bandwidth usage.

We consider epidemic-style information dissemination strategies that leverage the nonuniformity of host distribution over subnets (e.g., IP subnets) to optimize the information spread. Such epidemic-style strategies are based on random sampling of target hosts according to a sampling rule. In this paper, we consider the metric of total number of samplings (equivalently probes) to reach a given target fraction of the host population. We first identify the minimum number of samplings needed to reach a target fraction of hosts, assuming global information about the host distribution over subnets is available. We show that this optimum can be achieved either by a dynamic strategy, for which the sampling probabilities over subnets are allowed to vary over time, or, surprisingly, even by a static strategy, for which the sampling probabilities over subnets are fixed. These results provide insights about the best achievable performance and how different system parameters affect the number of sampling needed. We then consider simple online sampling strategies that do not require any prior knowledge of the distribution of hosts over subnets, but where each host biases sampling based on its observed sampling outcomes while keeping only O(1) state at any point in time. Using real data-sets from several large-scale Internet measurements, we evaluate significance of the system parameters that determine the sampling requirements and compare the performance of our proposed distribution-oblivious sampling strategies to the theoretical bound. Our results provide insights for the design of efficient information dissemination systems, as well as for the design of countermeasures against worms that use subnet-preferential scanning.

We consider the problem of traffic management in small networks with both wireless and wired devices, connected to the Internet through a single gateway. Examples of such networks are small office networks or residential networks, where typically traffic management is limited to flow prioritization through port-based filtering.

Dense, unmanaged 802.11 deployments tempt saboteurs into launching jamming attacks by injecting malicious interference. Nowadays, jammers can be portable devices that transmit intermittently at low power in order to conserve energy. In this paper, we first conduct extensive experiments on an indoor 802.11 network to assess the ability of two physical layer functions, rate adaptation and power control, in mitigating jamming. In the presence of a jammer we find that: (a) the use of popular rate adaptation algorithms can significantly degrade network performance and, (b) appropriate tuning of the carrier sensing threshold allows a transmitter to send packets even when being jammed and enables a receiver capture the desired signal. Based on our findings, we build ARES, an Anti-jamming REinforcement System, which tunes the parameters of rate adaptation and power control to improve the performance in the presence of jammers. ARES ensures that operations under benign conditions are unaffected. To demonstrate the effectiveness and generality of ARES, we evaluate it in three wireless testbeds: (a) an 802.11n WLAN with MIMO nodes, (b) an 802.11a/g mesh network with mobile jammers and (c) an 802.11a WLAN with TCP traffic. We observe that ARES improves the network throughput across all testbeds by up to 150%.

We consider wireless mesh networks, and exploit the inherent broadcast nature of wireless by making use of multipath routing. We present an optimization framework that enables us to derive optimal flow control, routing, scheduling, and rate adaptation schemes, where we use network coding to ease the routing problem. We prove optimality and derive a primal-dual algorithm that lays the basis for a practical protocol. We use simulation to show on realistic topologies that we can achieve 20-200% throughput improvement compared to single path routing, and several times compared to a recent related opportunistic protocol (MORE).

We consider epidemic-style information dissemination strategies that leverage the nonuniformity of host distribution over subnets (e.g., IP subnets) to optimize the information spread. Such epidemic-style strategies are based on random sampling of target hosts according to a sampling rule. The objective is to optimize the information spread with respect to minimizing the total number of samplings to reach a target fraction of the host population. This is of general interest for the design of efficient information dissemination systems and more specifically, to identify requirements for the containment of worms that use subnet preference scanning strategies. We first identify the optimum number of samplings to reach a target fraction of hosts, given global information about the host distribution over subnets. We show that the optimum can be achieved by either a dynamic strategy for which the per host sampling rate over subnets is allowed to vary over time or by a static strategy for which the sampling over subnets is fixed. These results appear to be novel and are informative about (a) what best possible performance is achievable and (b) what factors determine the performance gain over oblivious strategies such as uniform random scanning. We then consider several simple, online sampling strategies that require only local knowledge, where each host biases sampling based on its observed sampling outcomes and keeps only O(1) state at any point in time. Using real datasets from several large-scale Internet measurements, we evaluate the significance of the factors revealed by our analytical results on the sampling efficiency.

There has been extensive work on network architectures that support multi-path routing to improve performance in wireless mesh networks. However, previous work uses ad-hoc design principles that cannot guarantee any network-wide performance objectives such as conjointly maximizing resource utilization and improving fairness. In parallel, numerous theoretical results have addressed the issue of optimizing a combined metric of network utilization and fairness using techniques based on back-pressure scheduling, routing and flow control. However, the proposed theoretical algorithms are extremely difficult to implement in practice, especially in the presence of the 802.11 MAC and TCP. We propose Horizon, a novel system design for multi-path forwarding in wireless meshes, based on the theoretical results on back-pressure. Our design works with an unmodified TCP stack and on top of the existing 802.11 MAC. We modified the back-pressure approach to obtain a simple 802.11-compatible packet-forwarding heuristic and a novel, light-weight path estimator, while maintaining global optimality properties. We propose a delayed reordering algorithm that eliminates TCP timeouts while keeping TCP packet reordering to a minimum. We have evaluated our implementation on a 22-node testbed. We have shown that Horizon effectively utilizes available resources (disjoint paths). In contrast to previous work, our design not only avoids bottlenecks but also optimally load-balances traffic across them when needed, improving fairness among competing flows. To our knowledge, Horizon is the first practical wireless system based on back-pressure.

Designing high throughput wireless mesh networks involves solving interrelated scheduling, routing, and interference problems. In this paper, we exploit the broadcast properties and the path diversity of wireless meshes to implement an efficient multipath routing protocol, Multipath Code Casting (MC2).

Digital media companies have recently started embracing P2P networks as an alternative distribution mechanism. However, with current P2P swarming systems users need to download the full video, and hence have to wait a long time before they can start watching it. While a lot of effort has gone into optimizing the distribution of large files, little research has been done on enabling Video-on-Demand (VoD) functionality with P2P swarming systems. The main challenges reside in ensuring that users can start watching a movie at any point in time, with small start-up times and sustainable playback rates. In this work, we address the issues of providing VoD using P2P mesh-based networks. We investigate scheduling techniques, and network coding in particular. Using both simulations and a prototype implementation, we show that high-quality VoD is feasible, and give guidelines to build play-as-you-download P2P swarming systems with high playback rates and low start-up delays.

We focus on efficient protocols that enhance a network with topology awareness. We discuss centralized algorithms with provable performance, and introduce decentralized asynchronous heuristics that use only local information and local computations. These algorithms are based on distributed solutions of convex programs expressing optimization of various spectral properties of the matrix associated with the graph of the network topology. For example, these algorithms assign special weights to links crossing or directed towards small cuts by minimizing the second eigenvalue. Our main technical ingredient is to perform the decentralized asynchronous computations in a manner that preserves critical invariants of the exact second eigenvalue of the adjacency matrix associated with the network topology.

We consider the problem of broadcasting a live stream of data in an unstructured network. The broadcasting problem has been studied extensively for edge-capacitated networks. We give the first proof that whenever demand lambda + epsiv is feasible for epsiv > 0, a simple local-control algorithm is stable under demand lambda, and as a corollary a famous theorem of Edmonds. We then study the node-capacitated case and show a similar optimality result for the complete graph. We study through simulation the delay that users must wait in order to playback a video stream with a small number of skipped packets, and discuss the suitability of our algorithms for live video streaming.

Peer-to-peer technologies are increasingly becoming the medium of choice for delivering media content, both professional and homegrown, to large user populations. Indeed, current P2P swarming systems have been shown to be very efficient for large-scale content distribution with few server resources. However, such systems have been designed for generic file distribution and provide a limited user experience for viewing media content. For example, users need to wait to download the full video before they can start watching it. In general, the main challenge resides in designing systems that ensure that users can start watching a movie at any point in time, with small start-up times and sustainable playback rates. In this work, we address the issues of providing a Video-on-Demand (VoD) using P2P mesh-based networks. We show that providing high quality VoD using P2P is feasible using a combination of techniques including (a) network coding, (b) optimized resource allocation across different parts of the video, and (c) overlay topology management algorithms. Our evaluation also shows that systems that do not optimize in all these dimensions could significantly under-utilize the network resources resulting in poor VoD performance. We present our results based on simulations and a prototype implementation. Categories and Subject Descriptors

In this paper we present the first implementation of a P2P content distribution system that uses Network Coding. Using results from live trials with several hundred nodes, we provide a detailed performance analysis of such P2P system. In contrast to prior work, which mainly relies on monitoring P2P systems at particular locations, we are able to provide performance results from a variety of novel angles by monitoring all components in the P2P distribution.In particular, we show that Network Coding is practical in a P2P setting since it incurs little overhead, both in terms of CPU processing and IO activity, and it results in smooth, fast downloads, and efficient server utilization. We also study the importance of topology construction algorithms in real scenarios and study the effect of peers behind NATs and firewalls, showing that the system is surprisingly robust to large number of unreachable peers. Finally, we present performance results related to verifying network encoded blocks on-the-fly using special security primitives called Secure-Random-Checksums.

Peer-to-peer content distribution networks can suffer from malicious participants that corrupt content. Current systems verify blocks with traditional cryptographic signatures and hashes. However, these techniques do not apply well to more elegant schemes that use network coding techniques for efficient content distribution. Architectures that use network coding are prone to jamming attacks where the introduction of a few corrupted blocks can quickly result in a large number of bad blocks propagating through the system. Identifying such bogus blocks is difficult and requires the use of homomorphic hashing functions, which are computationally expensive. This paper presents a practical security scheme for network coding that reduces the cost of verifying blocks on-the-fly while efficiently preventing the propagation of malicious blocks. In our scheme, users not only cooperate to distribute the content, but (well-behaved) users also cooperate to protect themselves against malicious users by informing affected nodes when a malicious block is found. We analyze and study such cooperative security scheme and introduce elegant techniques to prevent DoS attacks. We show that the loss in the efficiency caused by the attackers is limited to the effort the attackers put to corrupt the communication, which is a natural lower bound in the damage of the system. We also show experimentally that checking as low as 1-5% of the received blocks is enough to guarantee low corruption rates.

Fast and effective distribution of software updates (a.k.a. patches) to millions of Internet users has evolved into a critical task over the last years. In this paper, we characterize "Windows Update", one of the largest update services in the world, with the aim to draw general guidelines on how to best design and architect a fast and effective planet-scale patch dissemination system. To this end, we analyze an extensive set of data traces collected over the period of a year, consisting of billions of queries from over 300 million computers. Based on empirical observations and analytical results, we identify interesting properties of today's update traffic and user behavior.Building on this analysis, we consider alternative patch delivery strategies such as caching and peer-to-peer and evaluate their performance. We identify key factors that determine the effectiveness of these schemes in reducing the server workload and the network traffic, and in speeding-up the patch delivery. Most of our findings are invariant properties induced by either user behavior or architectural characteristics of today's Internet, and thus apply to the general problem of Internet-wide dissemination of software updates.

We study hybrid search schemes for unstructured peer-to-peer networks. We quantify performance in terms of number of hits, network overhead, and response time. Our schemes combine flooding and random walks, look ahead and replication. We consider both regular topologies and topologies with supernodes. We introduce a general search scheme, of which flooding and random walks are special instances, and show how to use locally maintained network information to improve the performance of searching. Our main findings are: (a) a small number of supernodes in an otherwise regular topology can offer sharp savings in the performance of search, both in the case of search by flooding and search by random walk, particularly when it is combined with 1-step replication. We quantify, analytically and experimentally, that the reason of these savings is that the search is biased towards nodes that yield more information. (b) There is a generalization of search, of which flooding and random walk are special instances, which may take further advantage of locally maintained network information, and yield better performance than both flooding and random walk in clustered topologies. The method determines edge critically and is reminiscent of fundamental heuristics from the area of approximation algorithms.

We propose a new scheme for content distribution of large files that is based on network coding. With network coding, each node of the distribution network is able to generate and transmit encoded blocks of information. The randomization introduced by the coding process eases the scheduling of block propagation, and, thus, makes the distribution more efficient. This is particularly important in large unstructured overlay networks, where the nodes need to make block forwarding decisions based on local information only. We compare network coding to other schemes that transmit unencoded information (i.e. blocks of the original file) and, also, to schemes in which only the source is allowed to generate and transmit encoded packets. We study the performance of network coding in heterogeneous networks with dynamic node arrival and departure patterns, clustered topologies, and when incentive mechanisms to discourage free-riding are in place. We demonstrate through simulations of scenarios of practical interest that the expected file download time improves by more than 20-30% with network coding compared to coding at the server only and, by more than 2-3 times compared to sending unencoded information. Moreover, we show that network coding improves the robustness of the system and is able to smoothly handle extreme situations where the server and nodes leave the system.

We quantify the effectiveness of random walks for searching and construction of unstructured peer-to-peer (P2P) networks. We have identified two cases where the use of random walks for searching achieves better results than flooding: a) when the overlay topology is clustered, and h) when a client re-issues the same query while its horizon does not change much. For construction, we argue that an expander can he maintained dynamically with constant operations per addition. The key technical ingredient of our approach is a deep result of stochastic processes indicating that samples taken from consecutive steps of a random walk can achieve statistical properties similar to independent sampling (if the second eigenvalue of the transition matrix is hounded away from 1, which translates to good expansion of the network; such connectivity is desired, and believed to hold, in every reasonable network and network model). This property has been previously used in complexity theory for construction of pseudorandom number generators. We reveal another facet of this theory and translate savings in random bits to savings in processing overhead.

Spectral analysis of the Internet topology at the autonomous system (AS) level, by adapting the standard spectral filtering method of examining the eigenvectors corresponding to the largest eigenvalues of matrices related to the adjacency matrix of the topology is performed. We observe that the method suggests clusters of ASs with natural semantic proximity, such as geography or business interests. We examine how these clustering properties vary in the core and in the edge of the network, as well as across geographic areas, over time, and between real and synthetic data. We observe that these clustering properties may be suggestive of traffic patterns and thus have direct impact on the link stress of the network. Finally, we use the weights of the eigenvector corresponding to the first eigenvalue to obtain an alternative hierarchical ranking of the ASs.

It has been observed that the degrees of the topologies of several communication networks follow heavy tailed statistics. What is the impact of such heavy tailed statistics on the performance of basic communication tasks that a network is presumed to support? How does performance scale with the size of the network? We study routing in families of sparse random graphs whose degrees follow heavy tailed distributions. Instantiations of such random graphs have been proposed as models for the topology of the Internet at the level of Autonomous Systems as well as at the level of routers. Let n be the number of nodes. Suppose that for each pair of nodes with degrees du and dv we have O(dudv ) units of demand. Thus the total demand is O(n ). We argue analytically and experimentally that in the considered random graph model such demand patterns can be routed so that the flow through each link is at most O . This is to be compared with a bound # that holds for arbitrary graphs. Similar results were previously known for sparse random regular graphs, a.k.a. "expander graphs." The significance is that Internet-like topologies, which grow in a dynamic, decentralized fashion and appear highly inhomogeneous, can support routing with performance characteristics comparable to those of their regular counterparts, at least under the assumption of uniform demand and capacities. Our proof uses approximation algorithms for multicommodity flow and establishes strong bounds of a generalization of "expansion," namely "conductance." Besides routing, our bounds on conductance have further implications, most notably on the gap between first and second eigenvalues of the stochastic normalization of the adjacency matrix of the graph.