ACM SIGMETRICS Performance Evaluation Review

Published by Association for Computing Machinery
Online ISSN: 0163-5999
Publications
Conference Paper
Congestion control in TCP/AQM networks is expected to perform well for a wide-range of conditions, but recent advances in modeling and analysis indicate that present AQM (active queue management) schemes need an extra dose of adaptability to cope. The paper answers the call and proposes a self-tuning structure wherein AQM parameters are automatically tuned in response to on-line estimation of link capacity and traffic load. This approach is applicable to any AQM scheme that is parameterizable in terms of link capacity and TCP load. We describe this self-tuning structure, illustrate its application to PI (proportional-integral) and RED (random early detection) AQMs, provide stability analysis, and conduct ns simulations to compare with both fixed AQM schemes and the recently proposed adaptive RED.
 
Conference Paper
Fine-grained network measurement requires routers and switches to update large arrays of counters at very high link speed (e.g. 40 Gbps). A naive algorithm needs an infeasible amount of SRAM to store both the counters and a flow-to-counter association rule, so that arriving packets can update corresponding counters at link speed. This has made accurate per-flow measurement complex and expensive, and motivated approximate methods that detect and measure only the large flows. This paper revisits the problem of accurate per-flow measurement. We present a counter architecture, called Counter Braids, inspired by sparse random graph codes. In a nutshell, Counter Braids "compresses while counting". It solves the central problems (counter space and flow-to-counter association) of per-flow measurement by "braiding" a hierarchy of counters with random graphs. Braiding results in drastic space reduction by sharing counters among flows; and using random graphs generated on-the-fly with hash functions avoids the storage of flow-to-counter association. The Counter Braids architecture is optimal (albeit with a complex decoder) as it achieves the maximum compression rate asymptotically. For implementation, we present a low-complexity message passing decoding algorithm, which can recover flow sizes with essentially zero error. Evaluation on Internet traces demonstrates that almost all flow sizes are recovered exactly with only a few bits of counter space per flow.
 
Conference Paper
Storage device performance prediction is a key element of self-managed storage systems. The paper explores the application of a machine learning tool, CART (classification and regression trees) models, to storage device modeling. Our approach predicts a device's performance as a function of input workloads, requiring no knowledge of the device internals. We propose two uses of CART models: one that predicts per-request response times (and then derives aggregate values); one that predicts aggregate values directly from workload characteristics. After being trained on the device in question, both provide accurate black-box models across a range of test traces from real environments. Experiments show that these models predict the average and 90th percentile response time with a relative error as low as 19%, when the training workloads are similar to the testing workloads, and interpolate well across different workloads.
 
Article
Scientific codes are usually parallelized by partitioning a grid among processors. To achieve top performance it is necessary to partition the grid so as to balance workload and minimize communication/synchronization costs. This problem is particularly acute when the grid is irregular, changes over the course of the computation, and is not known until load time. Critical mapping and remapping decisions rest on the ability to accurately predict performance, given a description of a grid and its partition. This paper discusses one approach to this problem, and illustrates its use on a one-dimensional fluids code. The models constructed are shown to be accurate, and are used to find optimal remapping schedules.
 
Article
One of the key issues in providing end-to-end quality-of-service (QoS) guarantees in packet networks is how to determine a feasible path that satisfies a number of QoS constraints. For two or more additive constraints, the problem of finding a feasible path is NP-complete that cannot be exactly solved in polynomial time. Accordingly, several heuristics and approximation algorithms have been proposed for this problem. Many of these algorithms suffer from either excessive computational cost or low performance. In this paper, we provide an efficient approximation algorithm for finding a path subject to two additive constraints. The worst-case computational complexity of this algorithm is within a logarithmic number of calls to Dijkstra's shortest path algorithm. Its average complexity is even much lower than that, as demonstrated by simulation experiments. The performance of the proposed algorithm is justified via theoretical bounds that are provided for the optimal version of the path selection problem. To achieve further performance improvement, several extensions to the basic algorithm are also provided at very low computational cost. Extensive simulations are used to demonstrate the high performance of the proposed algorithm and to contrast it with other path selection algorithms.
 
Article
Most reliability analysis techniques and tools assume that a system is used for a mission consisting of a single phase. However, multiple phases are natural in many missions. The failure rates of components, system configuration, and success criteria may vary from phase to phase. In addition, the duration of a phase may be deterministic or random. Recently, several researchers have addressed the problem of reliability analysis of such systems using a variety of methods. A new technique for phased-mission system reliability analysis based on Boolean algebraic methods is described. Our technique is computationally efficient and is applicable to a large class of systems for which the failure criterion in each phase can be expressed as a fault tree (or an equivalent representation). Our technique avoids state space explosion that commonly plague Markov chain-based analysis. A phase algebra to account for the effects of variable configurations and success criteria from phase to phase was developed. Our technique yields exact (as opposed to approximate) results. The use of our technique was demonstrated by means of an example and present numerical results to show the effects of mission phases on the system reliability.
 
Lengths of paths from monitor apan-jp 
Doubletree redundancy, 95 th percentile. Inter-monitor redundancy on destinations, gross redundancy on router interfaces. 
Links and nodes coverage in comparison to classic probing 
Amount of probes sent 
Article
There is a growing interest in discovery of internet topology at the interface level. A new generation of highly distributed measurement systems is currently being deployed. Unfortunately, the research community has not examined the problem of how to perform such measurements efficiently and in a network-friendly manner. In this paper we make two contributions toward that end. First, we show that standard topology discovery methods (e.g., skitter) are quite inefficient, repeatedly probing the same interfaces. This is a concern, because when scaled up, such methods will generate so much traffic that they will begin to resemble DDoS attacks. We measure two kinds of redundancy in probing (intra- and inter-monitor) and show that both kinds are important. We show that straightforward approaches to addressing these two kinds of redundancy must take opposite tacks, and are thus fundamentally in conflict. Our second contribution is to propose and evaluate Doubletree, an algorithm that reduces both types of redundancy simultaneously on routers and end systems. The key ideas are to exploit the tree-like structure of routes to and from a single point in order to guide when to stop probing, and to probe each path by starting near its midpoint. Our results show that Doubletree can reduce both types of measurement load on the network dramatically, while permitting discovery of nearly the same set of nodes and links. We then show how to enable efficient communication between monitors through the use of Bloom filters.
 
TDD and FD capacity regions, and FD extension. The capacity region is plotted for equal maximum SNRs: γ max mb = γ max bm ≡ γ max bm/mb and two cases of maximum XINRs: (a) γ max bb = 1, γ max mm = 1 and (b) γ max bb = 1, γ max mm = 10. capacity region to the corresponding TDD capacity region. Let r b = log(1 + γ bm 1+γmm ), r m = log(1 + γ mb 1+γ bb ) denote DL and UL rates, respectively and let r max TDD,b = log(1 + γ max bm ), r max TDD,m = log(1 + γ max mb ) denote the maximum TDD rates. The FD capacity region is the set of all points (r b , r m ) such that P m ∈ [0, P m ], P b ∈ [0, P b ], while the TDD capacity region is the convex hull of the points (0, 0), (r max TDD,b , 0), and (0, r max TDD,m ). We also let s b = log(1 + γ max bm 1+γ max mm ) and 
Article
Full-duplex communication has the potential to substantially increase the throughput in wireless networks. However, the benefits of full-duplex are still not well understood. In this paper, we characterize the full-duplex rate gains in both single-channel and multi-channel use cases. For the single-channel case, we quantify the rate gain as a function of the remaining self-interference and SNR values. We also provide a sufficient condition under which the sum of uplink and downlink rates on a full-duplex channel is concave in the transmission power levels. Building on these results, we consider the multi-channel case. For that case, we introduce a new realistic model of a small form-factor (e.g., smartphone) full-duplex receiver and demonstrate its accuracy via measurements. We study the problem of jointly allocating power levels to different channels and selecting the frequency of maximum self-interference suppression, where the objective is maximizing the sum of the rates over uplink and downlink OFDM channels. We develop a polynomial time algorithm which is nearly optimal under very mild restrictions. To reduce the running time, we develop an efficient nearly-optimal algorithm under the high SINR approximation. Finally, we demonstrate via numerical evaluations the capacity gains in the different use cases and obtain insights into the impact of the remaining self-interference and wireless channel states on the performance.
 
A learning protocol R = (L, Y,
Article
Online services routinely mine user data to predict user preferences, make recommendations, and place targeted ads. Recent research has demonstrated that several private user attributes (such as political affiliation, sexual orientation, and gender) can be inferred from such data. Can a privacy-conscious user benefit from personalization while simultaneously protecting her private attributes? We study this question in the context of a rating prediction service based on matrix factorization. We construct a protocol of interactions between the service and users that has remarkable optimality properties: it is privacy-preserving, in that no inference algorithm can succeed in inferring a user's private attribute with a probability better than random guessing; it has maximal accuracy, in that no other privacy-preserving protocol improves rating prediction; and, finally, it involves a minimal disclosure, as the prediction accuracy strictly decreases when the service reveals less information. We extensively evaluate our protocol using several rating datasets, demonstrating that it successfully blocks the inference of gender, age and political affiliation, while incurring less than 5% decrease in the accuracy of rating prediction.
 
Macrostructure of any directed graph. 
Macrostructure of Twitter in July 2012. The size of the circles is proportional to the number of accounts in components. The labels on arrows give the number of arcs between components. 
Age of accounts in each component. CCDF of accounts in a component according to the account creation date. 
The distribution of number of tweets by component. Accounts with no tweets are filtered out (see Table 3). 
Article
Twitter is one of the largest social networks using exclusively directed links among accounts. This makes the Twitter social graph much closer to the social graph supporting real life communications than, for instance, Facebook. Therefore, understanding the structure of the Twitter social graph is interesting not only for computer scientists, but also for researchers in other fields, such as sociologists. However, little is known about how the information propagation in Twitter is constrained by its inner structure. In this paper, we present an in-depth study of the macroscopic structure of the Twitter social graph unveiling the highways on which tweets propagate, the specific user activity associated with each component of this macroscopic structure, and the evolution of this macroscopic structure with time for the past 6 years. For this study, we crawled Twitter to retrieve all accounts and all social relationships (follow links) among accounts; the crawl completed in July 2012 with 505 million accounts interconnected by 23 billion links. Then, we present a methodology to unveil the macroscopic structure of the Twitter social graph. This macroscopic structure consists of 8 components defined by their connectivity characteristics. Each component group users with a specific usage of Twitter. For instance, we identified components gathering together spammers, or celebrities. Finally, we present a method to approximate the macroscopic structure of the Twitter social graph in the past, validate this method using old datasets, and discuss the evolution of the macroscopic structure of the Twitter social graph during the past 6 years.
 
Article
We consider a system of parallel queues where tasks are assigned (dispatched) to one of the available servers upon arrival. The dispatching decision is based on the full state information, i.e., on the sizes of the new and existing jobs. We are interested in minimizing the so-called mean slowdown criterion corresponding to the mean of the sojourn time divided by the processing time. Assuming no new jobs arrive, the shortest-processing-time-product (SPTP) schedule is known to minimize the slowdown of the existing jobs. The main contribution of this paper is three-fold: 1) To show the optimality of SPTP with respect to slowdown in a single server queue under Poisson arrivals; 2) to derive the so-called size-aware value functions for M/G/1-FIFO/LIFO/SPTP/SPT/SRPT with general holding costs of which the slowdown criterion is a special case; and 3) to utilize the value functions to derive efficient dispatching policies so as to minimize the mean slowdown in a heterogeneous server system. The derived policies offer a significantly better performance than e.g., the size-aware-task-assignment with equal load (SITA-E) and least-work-left (LWL) policies.
 
Article
Recent studies show that a large fraction of Internet traffic is originated by Content Providers (CPs) such as content distribution networks and hyper-giants. To cope with the increasing demand for content, CPs deploy massively distributed server infrastructures. Thus, content is available in many network locations and can be downloaded by traversing different paths in a network. Despite the prominent server location and path diversity, the decisions on how to map users to servers by CPs and how to perform traffic engineering by ISPs, are independent. This leads to a lose-lose situation as CPs are not aware about the network bottlenecks nor the location of end-users, and the ISPs struggle to cope with rapid traffic shifts caused by the dynamic CP server selection process. In this paper we propose and evaluate Content-aware Traffic Engineering (CaTE), which dynamically adapts the traffic demand for content hosted on CPs by utilizing ISP network information and end-user location during the server selection process. This leads to a win-win situation because CPs are able to enhance their end-user to server mapping and ISPs gain the ability to partially influence the traffic demands in their networks. Indeed, our results using traces from a Tier-1 ISP show that a number of network metrics can be improved when utilizing CaTE.
 
Toy Example-with link weights 
Simplified Model 
Traffic shifts from one shortest path to another 
CDFs of Umax over all TMs for BGPawareLWO and IntraLWO-resulting for the updated topology 
Article
A well-known approach to intradomain traffic engineering consists in finding the set of link weights that minimizes a network-wide objective function for a given intradomain traffic matrix. This approach is inadequate because it ignores a potential impact on interdomain routing due to hot-potato routing policies. This may result in changes in the intradomain traffic matrix that have not been anticipated by the link weights optimizer, possibly leading to degraded network performance. We propose a BGP-aware link weights optimization method that takes these hot-potato effects into account. This method uses the interdomain traffic matrix and other available BGP data, to extend the intradomain topology with external virtual nodes and links, on which all the well-tuned heuristics of a classical link weights optimizer can be applied. Our method can also optimize the traffic on the interdomain peering links.
 
In multiple-queue backfilling, initial partition boundaries adapt as workload conditions change. In this example, we consider 30 processors and three partitions (queues). 
Conference Paper
We describe a new, non-FCFS policy to schedule parallel jobs on systems that may be part of a computationalgrid . Our algorithm continuously monitors the system (i.e., the intensity of incoming jobs and variability of their resource demands), and adapts its scheduling parameters according to workload fluctuations. The proposed policy is based on backfilling, which reduces resource fragmentation by executing jobs in an order different than their arrivalwit hout delaying certain previously submitted jobs. We maintain multiple job queues that effectively separate jobs according to their projected execution time. Our policy supports different job priorities and job reservations, making it appropriate for scheduling jobs on parallel systems that are part of a computational grid. Detailed performance comparisons via simulation using traces from the Parallel Workload Archive indicate that the proposed policy consistently outperforms traditional backfilling.
 
Article
Graph sampling via crawling has been actively considered as a generic and important tool for collecting uniform node samples so as to consistently estimate and uncover various characteristics of complex networks. The so-called simple random walk with re-weighting (SRW-rw) and Metropolis-Hastings (MH) algorithm have been popular in the literature for such unbiased graph sampling. However, an unavoidable downside of their core random walks -- slow diffusion over the space, can cause poor estimation accuracy. In this paper, we propose non-backtracking random walk with re-weighting (NBRW-rw) and MH algorithm with delayed acceptance (MHDA) which are theoretically guaranteed to achieve, at almost no additional cost, not only unbiased graph sampling but also higher efficiency (smaller asymptotic variance of the resulting unbiased estimators) than the SRW-rw and the MH algorithm, respectively. In particular, a remarkable feature of the MHDA is its applicability for any non-uniform node sampling like the MH algorithm, but ensuring better sampling efficiency than the MH algorithm. We also provide simulation results to confirm our theoretical findings.
 
Nodes in ISP A, | V ( A, T ) | , vs. total torrent size, | V ( T ) | , for US1 (top) and EU1 (bottom). 
Article
A substantial amount of work has recently gone into localizing BitTorrent traffic within an ISP in order to avoid excessive and often times unnecessary transit costs. Several architectures and systems have been proposed and the initial results from specific ISPs and a few torrents have been encouraging. In this work we attempt to deepen and scale our understanding of locality and its potential. Looking at specific ISPs, we consider tens of thousands of concurrent torrents, and thus capture ISP-wide implications that cannot be appreciated by looking at only a handful of torrents. Secondly, we go beyond individual case studies and present results for the top 100 ISPs in terms of number of users represented in our dataset of up to 40K torrents involving more than 3.9M concurrent peers and more than 20M in the course of a day spread in 11K ASes. We develop scalable methodologies that permit us to process this huge dataset and answer questions such as: "\emph{what is the minimum and the maximum transit traffic reduction across hundreds of ISPs?}", "\emph{what are the win-win boundaries for ISPs and their users?}", "\emph{what is the maximum amount of transit traffic that can be localized without requiring fine-grained control of inter-AS overlay connections?}", "\emph{what is the impact to transit traffic from upgrades of residential broadband speeds?}".
 
Time duration that peers unchoked each other via a regular unchoke, averaged over all runs. Darker squares represent longer unchoke times (the unit of the color bar on the right is in seconds). Peers 1 to 13 have a 20 kB/s upload limit, peers 14 to 27 have a 50 kB/s upload limit, and peers 28 to 40 have a 200 kB/s upload limit. The seed (peer 41) is limited to 200 kB/s. The creation of clusters is clearly visible.
Total number of bytes uploaded by peers to each other, averaged over all runs. Darker squares represent more data (the unit of the color bar on the right is in bytes). Peers 1 to 13 have a 20 kB/s upload limit, peers 14 to 27 have a 50 kB/s upload limit, and peers 28 to 40 have a 200 kB/s upload limit. The seed (peer 41) is limited to 200 kB/s. Fast peers upload much more data than the rest.
Scatterplot of peers' upload utilization for all 60-second time intervals during the download, in the presence of a wellprovisioned seed (limited to 200 kB/s). Each point represents the average upload utilization over all peers for a given experiment run. Utilization is kept high during most of the download session.
Duration of all unchokes (regular and optimistic) performed by a well-provisioned seed to each peer. Results for a single representative run. Peers 1 to 13 have a 20 kB/s upload limit, peers 14 to 27 have a 50 kB/s upload limit, and peers 28 to 40 have a 200 kB/s upload limit. The seed (peer 41) provides uniform service to all leechers.
Time duration that peers unchoked each other via a regular unchoke, averaged over all runs. Darker squares represent longer unchoke times (the unit of the color bar on the right is in seconds). Peers 1 to 12 have a 20 kB/s upload limit, peers 13 to 26 have a 50 kB/s upload limit, and peers 28 to 40 have a 200 kB/s upload limit. The seed (peer 27) is limited to 100 kB/s. There is no discernible clustering.
Article
Peer-to-peer protocols play an increasingly instrumental role in Internet content distribution. It is therefore important to gain a complete understanding of how these protocols behave in practice and how their operating parameters affect overall system performance. This paper presents the first detailed experimental investigation of the peer selection strategy in the popular BitTorrent protocol. By observing more than 40 nodes in instrumented private torrents, we validate three protocol properties that, though believed to hold, have not been previously demonstrated experimentally: the clustering of similar-bandwidth peers, the effectiveness of BitTorrent's sharing incentives, and the peers' high uplink utilization. In addition, we observe that BitTorrent's modified choking algorithm in seed state provides uniform service to all peers, and that an underprovisioned initial seed leads to absence of peer clustering and less effective sharing incentives. Based on our results, we provide guidelines for seed provisioning by content providers, and discuss a tracker protocol extension that addresses an identified limitation of the protocol.
 
A Markov-modulated process for the aggregation of n homogeneous MMOOs
Article
The practicality of the stochastic network calculus (SNC) is often questioned on grounds of potential looseness of its performance bounds. In this paper it is uncovered that for bursty arrival processes (specifically Markov-Modulated On-Off (MMOO)), whose amenability to \textit{per-flow} analysis is typically proclaimed as a highlight of SNC, the bounds can unfortunately indeed be very loose (e.g., by several orders of magnitude off). In response to this uncovered weakness of SNC, the (Standard) per-flow bounds are herein improved by deriving a general sample-path bound, using martingale based techniques, which accommodates FIFO, SP, EDF, and GPS scheduling. The obtained (Martingale) bounds gain an exponential decay factor of ${\mathcal{O}}(e^{-\alpha n})$ in the number of flows $n$. Moreover, numerical comparisons against simulations show that the Martingale bounds are remarkably accurate for FIFO, SP, and EDF scheduling; for GPS scheduling, although the Martingale bounds substantially improve the Standard bounds, they are numerically loose, demanding for improvements in the core SNC analysis of GPS.
 
Article
The non-preemptive priority queueing with a finite buffer is considered. We introduce a randomized push-out buffer management mechanism which allows to control very efficiently the loss probability of priority packets. The packet loss probabilities for priority and non-priority traffic are calculated using the generating function approach. In the particular case of the standard non-randomized push-out scheme we obtain explicit analytic expressions. The theoretical results are illustrated by numerical examples. The randomized push-out scheme is compared with the threshold based push-out scheme. It turns out that the former is much easier to tune than the latter. The proposed scheme can be applied to the Differentiated Services of the Internet.
 
Article
TTL caching models have recently regained significant research interest, largely due to their ability to fit popular caching policies such as LRU. This paper advances the state-of-the-art analysis of TTL-based cache networks by developing two exact methods with orthogonal generality and computational complexity. The first method generalizes existing results for line networks under renewal requests to the broad class of caching policies whereby evictions are driven by stopping times. The obtained results are further generalized, using the second method, to feedforward networks with Markov arrival processes (MAP) requests. MAPs are particularly suitable for non-line networks because they are closed not only under superposition and splitting, as known, but also under input-output caching operations as proven herein for phase-type TTL distributions. The crucial benefit of the two closure properties is that they jointly enable the first exact analysis of feedforward networks of TTL caches in great generality.
 
Article
In this note, we present preliminary results on the use of "network calculus" for parallel processing systems, specifically MapReduce. We also numerically evaluate the "generalized" (strong) stochastic burstiness bound based on publicly posted data describing an actual MapReduce workload of a Facebook datacenter.
 
Article
Recent studies on AS-level Internet connectivity have attracted considerable attention. These studies have exclusively relied on BGP data from the Oregon route-views [University of Oregon Route Views Project, http://www.routeviews.org] to derive some unexpected and intriguing results. The Oregon route-views data sets reflect AS peering relationships, as reported by BGP, seen from a handful of vantage points in the global Internet. The possibility that these data sets may provide only a very sketchy picture of the complete inter-AS connectivity of the Internet has received little scrutiny. By augmenting the Oregon route-views data with BGP summary information from a large number of Internet Looking Glass sites and with routing policy information from Internet Routing Registry (IRR) databases, we find that (1) a significant number of existing AS peering relationships remain hidden from most BGP routing tables, (2) the AS peering relationships with tier-1 ASs are in general more easily observed than those with non-tier-1 ASs, and (3) there are at least about 40% more AS peering relationships in the Internet than commonly-used BGP-derived AS maps reveal (but only about 4% more ASs). These findings point out the need for continuously questioning the applicability and completeness of data sets at hand when establishing the generality of any particular Internet-specific observation and for assessing its (in)sensitivity to deficiencies in the measurements.
 
Article
Data centers have emerged as promising resources for demand response, particularly for emergency demand response (EDR), which saves the power grid from incurring blackouts during emergency situations. However, currently, data centers typically participate in EDR by turning on backup (diesel) generators, which is both expensive and environmentally unfriendly. In this paper, we focus on "greening" demand response in multi-tenant data centers, i.e., colocation data centers, by designing a pricing mechanism through which the data center operator can efficiently extract load reductions from tenants during emergency periods to fulfill energy reduction requirement for EDR. In particular, we propose a pricing mechanism for both mandatory and voluntary EDR programs, ColoEDR, that is based on parameterized supply function bidding and provides provably near-optimal efficiency guarantees, both when tenants are price-taking and when they are price-anticipating. In addition to analytic results, we extend the literature on supply function mechanism design, and evaluate ColoEDR using trace-based simulation studies. These validate the efficiency analysis and conclude that the pricing mechanism is both beneficial to the environment and to the data center operator (by decreasing the need for backup diesel generation), while also aiding tenants (by providing payments for load reductions).
 
Functioning of (a) an optimal scheduling policy and (b) a probabilistic scheduling policy.
Our Tahoe testbed with average ping (RTT) and bandwidth measurements among three data centers in New Jersey, Texas, and California 
Article
Modern distributed storage systems offer large capacity to satisfy the exponentially increasing need of storage space. They often use erasure codes to protect against disk and node failures to increase reliability, while trying to meet the latency requirements of the applications and clients. This paper provides an insightful upper bound on the average service delay of such erasure-coded storage with arbitrary service time distribution and consisting of multiple files. Not only does the result supersede known delay bounds that only work for a single file, it also enables a novel problem of joint latency and storage cost minimization over three dimensions: selecting the erasure code, placement of encoded chunks, and optimizing scheduling policy. The problem is efficiently solved via the computation of a sequence of convex approximations with provable convergence. We further prototype our solution in an open-source, cloud storage deployment over three geographically distributed data centers. Experimental results validate our theoretical delay analysis and show significant latency reduction, providing valuable insights into the proposed latency-cost tradeoff in erasure-coded storage.
 
Average hourly spot market price during the week of 01/01/200501/07/2005 for LA1 Zone [6].
Block diagram for the basic model.
Periodic W (t) process in the example.
Block diagram for the extended model with delay tolerant and delay intolerant workloads.
Article
Since the electricity bill of a data center constitutes a significant portion of its overall operational costs, reducing this has become important. We investigate cost reduction opportunities that arise by the use of uninterrupted power supply (UPS) units as energy storage devices. This represents a deviation from the usual use of these devices as mere transitional fail-over mechanisms between utility and captive sources such as diesel generators. We consider the problem of opportunistically using these devices to reduce the time average electric utility bill in a data center. Using the technique of Lyapunov optimization, we develop an online control algorithm that can optimally exploit these devices to minimize the time average cost. This algorithm operates without any knowledge of the statistics of the workload or electricity cost processes, making it attractive in the presence of workload and pricing uncertainties. An interesting feature of our algorithm is that its deviation from optimality reduces as the storage capacity is increased. Our work opens up a new area in data center power management.
 
Article
Recently several CSMA algorithms based on the Glauber dynamics model have been proposed for multihop wireless scheduling, as viable solutions to achieve the throughput optimality, yet are simple to implement. However, their delay performances still remain unsatisfactory, mainly due to the nature of the underlying Markov chains that imposes a fundamental constraint on how the link state can evolve over time. In this paper, we propose a new approach toward better queueing and delay performance, based on our observation that the algorithm needs not be Markovian, as long as it can be implemented in a distributed manner, achieve the same throughput optimality, while offering far better delay performance for general network topologies. Our approach hinges upon utilizing past state information observed by local link and then constructing a high-order Markov chain for the evolution of the feasible link schedules. We show in theory and simulation that our proposed algorithm, named delayed CSMA, adds virtually no additional overhead onto the existing CSMA-based algorithms, achieves the throughput optimality under the usual choice of link weight as a function of local queue length, and also provides much better delay performance by effectively `de-correlating' the link state process (thus removing link starvation) under any arbitrary network topology. From our extensive simulations we observe that the delay under our algorithm can be often reduced by a factor of 20 over a wide range of scenarios, compared to the standard Glauber-dynamics-based CSMA algorithm.
 
Characteristics of random walks on a K3 edge Markovian graph model.
Article
In this paper we study the behavior of a continuous time random walk (CTRW) on a stationary and ergodic time varying dynamic graph. We establish conditions under which the CTRW is a stationary and ergodic process. In general, the stationary distribution of the walker depends on the walker rate and is difficult to characterize. However, we characterize the stationary distribution in the following cases: i) the walker rate is significantly larger or smaller than the rate in which the graph changes (time-scale separation), ii) the walker rate is proportional to the degree of the node that it resides on (coupled dynamics), and iii) the degrees of node belonging to the same connected component are identical (structural constraints). We provide examples that illustrate our theoretical findings.
 
Article
Motivated by emerging big streaming data processing paradigms (e.g., Twitter Storm, Streaming MapReduce), we investigate the problem of scheduling graphs over a large cluster of servers. Each graph is a job, where nodes represent compute tasks and edges indicate data-flows between these compute tasks. Jobs (graphs) arrive randomly over time, and upon completion, leave the system. When a job arrives, the scheduler needs to partition the graph and distribute it over the servers to satisfy load balancing and cost considerations. Specifically, neighboring compute tasks in the graph that are mapped to different servers incur load on the network; thus a mapping of the jobs among the servers incurs a cost that is proportional to the number of "broken edges". We propose a low complexity randomized scheduling algorithm that, without service preemptions, stabilizes the system with graph arrivals/departures; more importantly, it allows a smooth trade-off between minimizing average partitioning cost and average queue lengths. Interestingly, to avoid service preemptions, our approach does not rely on a Gibbs sampler; instead, we show that the corresponding limiting invariant measure has an interpretation stemming from a loss system.
 
Normalized monthly premium for a single period over five years
Insurance prices
Article
When a company migrates to cloud storage, the way back is neither fast nor cheap. The company is then locked up in the storage contract and exposed to upward market prices, which reduce the company’s profit and may even bring it below zero. We propose a protection means based on an insurance contract, by which the cloud purchaser is indem- nified when the current storage price exceeds a pre-defined threshold. By applying the financial options theory, we pro- vide a formula for the insurance price (the premium). By using historical data on market prices for disks, we apply the formula in realistic scenarios. We show that the pre- mium grows nearly quadratically with the duration of the coverage period as long as this is below one year, but grows more slowly, though faster than linearly, over longer cover- age periods.
 
Cosine of the angle between f (θ u A) and the row space of A for various b. 
Article
Given a set of pairwise comparisons, the classical ranking problem computes a single ranking that best represents the preferences of all users. In this paper, we study the problem of inferring individual preferences, arising in the context of making personalized recommendations. In particular, we assume that there are $n$ users of $r$ types; users of the same type provide similar pairwise comparisons for $m$ items according to the Bradley-Terry model. We propose an efficient algorithm that accurately estimates the individual preferences for almost all users, if there are $r \max \{m, n\}\log m \log^2 n$ pairwise comparisons per type, which is near optimal in sample complexity when $r$ only grows logarithmically with $m$ or $n$. Our algorithm has three steps: %first, for each user, project its $\binom{m}{2}$-dimensional vector of pairwise comparisons %onto an $m$-dimensional linear subspace to get the \nb{so-called: Bruce suggests delete it and change the whole sentence to: first, for each user, compute the \emph{net-win} vector which is a projection of its $\binom{m}{2}$-dimensional vector of pairwise comparisons onto an $m$-dimensional linear subspace; second, cluster the users based on the net-win vectors; third, estimate a single preference for each cluster separately. The net-win vectors are much less noisy than the high dimensional vectors of pairwise comparisons and clustering is more accurate after the projection as confirmed by numerical experiments. Moreover, we show that, when a cluster is only approximately correct, the maximum likelihood estimation for the Bradley-Terry model is still close to the true preference.
 
Article
Microgrids represent an emerging paradigm of future electric power systems that can utilize both distributed and centralized generations. Two recent trends in microgrids are the integration of local renewable energy sources (such as wind farms) and the use of co-generation (i.e., to supply both electricity and heat). However, these trends also bring unprecedented challenges to the design of intelligent control strategies for microgrids. Traditional generation scheduling paradigms rely on perfect prediction of future electricity supply and demand. They are no longer applicable to microgrids with unpredictable renewable energy supply and with co-generation (that needs to consider both electricity and heat demand). In this paper, we study online algorithms for the microgrid generation scheduling problem with intermittent renewable energy sources and co-generation, with the goal of maximizing the cost-savings with local generation. Based on the insights from the structure of the offline optimal solution, we propose a class of competitive online algorithms, called CHASE (Competitive Heuristic Algorithm for Scheduling Energy-generation), that track the offline optimal in an online fashion. Under typical settings, we show that CHASE achieves the best competitive ratio among all deterministic online algorithms, and the ratio is no larger than a small constant 3.
 
Article
There has appeared in the literature a great number of metrics that attempt to measure the effort or complexity in developing and understanding software(1). There have also been several attempts to independently validate these measures on data from different organizations gathered by different people(2). These metrics have many purposes. They can be used to evaluate the software development process or the software product. They can be used to estimate the cost and quality of the product. They can also be used during development and evolution of the software to monitor the stability and quality of the product. Among the most popular metrics have been the software science metrics of Halstead, and the cyclomatic complexity metric of McCabe. One question is whether these metrics actually measure such things as effort and complexity. One measure of effort may be the time required to produce a product. One measure of complexity might be the number of errors made during the development of a product. A second question is how these metrics compare with standard size measures, such as the number of source lines or the number of executable statements, i.e., do they do a better job of predicting the effort or the number of errors? Lastly, how do these metrics relate to each other?
 
Article
Since Tassiulas and Ephremides proposed the maximum weight scheduling algorithm of throughput-optimality for constrained queueing networks in 1992, extensive research efforts have been made for resolving its high complexity issue under various directions. In this paper, we resolve the issue by developing a generic framework for designing throughput-optimal and low-complexity scheduling algorithms. Under the framework, an algorithm updates current schedules via an interaction with a given oracle system that can generate a solution of a certain discrete optimization problem in a finite number of interactive queries. Therefore, one can design a variety of scheduling algorithms under this framework by choosing different oracles, e.g., the exhaustive search (ES), the markov chain monte carlo (MCMC), the belief propagation (BP) and the cutting-plane (CP) algorithms. The complexity of the resulting algorithm is decided by the number of operations required for an oracle processing a single query, which is typically very small. Somewhat surprisingly, we prove that an algorithm using any such oracle is throughput-optimal for general constrained queueing network models that arise in the context of emerging large-scale communication networks. In particular, the `pick-and-compare' algorithms developed by Tassiulas in 1998 and recently developed queue-based CSMA algorithms can be also understood as special cases of such algorithms using ES and MCMC oracles, respectively. To our best knowledge, our result is the first that establishes a rigorous connection between iterative optimization methods and low-complexity scheduling algorithms, which we believe provides various future directions and new insights in both areas.
 
Article
One typical use case of large-scale distributed computing in data centers is to decompose a computation job into many independent tasks and run them in parallel on different machines, sometimes known as the "embarrassingly parallel" computation. For this type of computation, one challenge is that the time to execute a task for each machine is inherently variable, and the overall response time is constrained by the execution time of the slowest machine. To address this issue, system designers introduce task replication, which sends the same task to multiple machines, and obtains result from the machine that finishes first. While task replication reduces response time, it usually increases resource usage. In this work, we propose a theoretical framework to analyze the trade-off between response time and resource usage. We show that, while in general, there is a tension between response time and resource usage, there exist scenarios where replicating tasks judiciously reduces completion time and resource usage simultaneously. Given the execution time distribution for machines, we investigate the conditions for a scheduling policy to achieve optimal performance trade-off, and propose efficient algorithms to search for optimal or near-optimal scheduling policies. Our analysis gives insights on when and why replication helps, which can be used to guide scheduler design in large-scale distributed computing systems.
 
Article
Our model is a constrained homogeneous random walk in a nonnegative orthant Z_+^d. The convergence to stationarity for such a random walk can often be checked by constructing a Lyapunov function. The same Lyapunov function can also be used for computing approximately the stationary distribution of this random walk, using methods developed by Meyn and Tweedie. In this paper we show that, for this type of random walks, computing the stationary probability exactly is an undecidable problem: no algorithm can exist to achieve this task. We then prove that computing large deviation rates for this model is also an undecidable problem. We extend these results to a certain type of queueing systems. The implication of these results is that no useful formulas for computing stationary probabilities and large deviations rates can exist in these systems.
 
Article
Users expect fast and fluid response from today's cloud infrastructure. Large-scale computing frameworks such as MapReduce divide jobs into many parallel tasks and execute them on different machines to enable faster processing. But the tasks on the slowest machines (straggling tasks) become the bottleneck in the completion of the job. One way to combat the variability in machine response time, is to add replicas of straggling tasks and wait for one copy to finish. In this paper we analyze how task replication strategies can be used to reduce latency, and their impact on the cost of computing resources. We use extreme value theory to show that the tail of the execution time distribution is the key factor in characterizing the trade-off between latency and computing cost. From this trade-off we can determine which task replication strategies reduce latency, without a large increase in computing cost. We also propose a heuristic algorithm to search for the best replication strategies when it is difficult to fit a simple distribution to model the empirical behavior of task execution time, and use the proposed analysis techniques. Evaluation of the heuristic policies on Google Trace data shows a significant latency reduction compared to the replication strategy used in MapReduce.
 
Article
Switched queueing networks model wireless networks, input queued switches and numerous other networked communications systems. For single-hop networks, we consider a (α,g)-switch policy} which combines the MaxWeight policies with bandwidth sharing networks -- a further well studied model of Internet congestion. We prove the maximum stability property for this class of randomized policies. Thus these policies have the same first order behavior as the MaxWeight policies. However, for multihop networks some of these generalized polices address a number of critical weakness of the MaxWeight/BackPressure policies. For multihop networks with fixed routing, we consider the Proportional Scheduler (or (1,log)-policy). In this setting, the BackPressure policy is maximum stable, but must maintain a queue for every route-destination, which typically grows rapidly with a network's size. However, this proportionally fair policy only needs to maintain a queue for each outgoing link, which is typically bounded in number. As is common with Internet routing, by maintaining per-link queueing each node only needs to know the next hop for each packet and not its entire route. Further, in contrast to BackPressure, the Proportional Scheduler does not compare downstream queue lengths to determine weights, only local link information is required. This leads to greater potential for decomposed implementations of the policy. Through a reduction argument and an entropy argument, we demonstrate that, whilst maintaining substantially less queueing overhead, the Proportional Scheduler achieves maximum throughput stability.
 
Article
We consider a large-scale service system model motivated by the problem of efficient placement of virtual machines to physical host machines in a network cloud, so that the total number of occupied hosts is minimized. Customers of different types arrive to a system with an infinite number of servers. A server packing configuration is the vector $k = (k_i)$, where $k_i$ is the number of type-$i$ customers that the server "contains". Packing constraints are described by a fixed finite set of allowed configurations. Upon arrival, each customer is placed into a server immediately, subject to the packing constraints; the server can be idle or already serving other customers. After service completion, each customer leaves its server and the system. It was shown recently that a simple real-time algorithm, called Greedy, is asymptotically optimal in the sense of minimizing $\sum_k X_k^{1+\alpha}$ in the stationary regime, as the customer arrival rates grow to infinity. (Here \alpha >0, and $X_k$ denotes the number of servers with configuration $k$.) In particular, when parameter \alpha is small, Greedy approximately solves the problem of minimizing $\sum_k X_k$, the number of occupied hosts. In this paper we introduce the algorithm called Greedy with sublinear Safety Stocks (GSS), and show that it asymptotically solves the exact problem of minimizing $\sum_k X_k$. An important feature of the algorithm is that sublinear safety stocks of $X_k$ are created automatically - when and where necessary - without having to determine a priori where they are required. Moreover, we also provide a tight characterization of the rate of convergence to optimality under GSS. The GSS algorithm is as simple as Greedy, and uses no more system state information than Greedy does.
 
An illustration of the allocation for N = 2
Article
Social utility maximization refers to the process of allocating resources in such a way that the sum of agents' utilities is maximized under the system constraints. Such allocation arises in several problems in the general area of communications, including unicast (and multicast multi-rate) service on the Internet, as well as in applications with (local) public goods, such as power allocation in wireless networks, spectrum allocation, etc. Mechanisms that implement such allocations in Nash equilibrium have also been studied but either they do not possess full implementation property, or are given in a case-by-case fashion, thus obscuring fundamental understanding of these problems. In this paper we propose a unified methodology for creating mechanisms that fully implement, in Nash equilibria, social utility maximizing functions arising in various contexts where the constraints are convex. The construction of the mechanism is done in a systematic way by considering the dual optimization problem. In addition to the required properties of efficiency and individual rationality that such mechanisms ought to satisfy, three additional design goals are the focus of this paper: a) the size of the message space scaling linearly with the number of agents (even if agents' types are entire valuation functions), b) allocation being feasible on and off equilibrium, and c) strong budget balance at equilibrium and also off equilibrium whenever demand is feasible.
 
Article
Making use of predictions is a crucial, but under-explored, area of online algorithms. This paper studies a class of online optimization problems where we have external noisy predictions available. We propose a stochastic prediction error model that generalizes prior models in the learning and stochastic control communities, incorporates correlation among prediction errors, and captures the fact that predictions improve as time passes. We prove that achieving sublinear regret and constant competitive ratio for online algorithms requires the use of an unbounded prediction window in adversarial settings, but that under more realistic stochastic prediction error models it is possible to use Averaging Fixed Horizon Control (AFHC) to simultaneously achieve sublinear regret and constant competitive ratio in expectation using only a constant-sized prediction window. Furthermore, we show that the performance of AFHC is tightly concentrated around its mean.
 
Article
Previous scalable protocols for downloading large, popular files from a single server include batching and cyclic multicast. With batching, clients wait to begin receiving a requested file until the beginning of its next multicast transmission, which collectively serves all of the waiting clients that have accumulated up to that point. With cyclic multicast, the file data is cyclically transmitted on a multicast channel. Clients can begin listening to the channel at an arbitrary point in time, and continue listening until all of the file data has been received.This paper first develops lower bounds on the average and maximum client delay for completely downloading a file, as functions of the average server bandwidth used to serve requests for that file, for systems with homogeneous clients. The results show that neither cyclic multicast nor batching consistently yields performance close to optimal. New hybrid download protocols are proposed that achieve within 15% of the optimal maximum delay and 20% of the optimal average delay in homogeneous systems.For heterogeneous systems in which clients have widely varying achievable reception rates, an additional design question concerns the use of high rate transmissions, which can decrease delay for clients that can receive at such rates, in addition to low rate transmissions that can be received by all clients. A new scalable download protocol for such systems is proposed, and its performance is compared to that of alternative protocols as well as to new lower bounds on maximum client delay. The new protocol achieves within 25% of the optimal maximum client delay in all scenarios considered.
 
Number of rounds in the test, Q(m) affects the number of ads that can be detected at least 8 and 15 rounds required for T ′ (t) and T (t) respectively.
The performance improves with number of collaborating users n.
As the size of the ad-pool A increases, the (personalized) ads become similar to effective recommendations, making it hard to detect (Type II error is large).
Type II error rate decreases as bias probability γ increases.
Variation of Type I + Type II error rates with perturbations in the algorithm's estimate of the average number of effective items˜fitems˜ items˜f ([m]). Data set : D3, Algorithm : L1 + A1, A = 8, γ = 0.4, n = 100.
Article
With a vast number of items, web-pages, and news to choose from, online services and the customers both benefit tremendously from personalized recommender systems. Such systems however provide great opportunities for targeted advertisements, by displaying ads alongside genuine recommendations. We consider a biased recommendation system where such ads are displayed without any tags (disguised as genuine recommendations), rendering them indistinguishable to a single user. We ask whether it is possible for a small subset of collaborating users to detect such a bias. We propose an algorithm that can detect such a bias through statistical analysis on the collaborating users' feedback. The algorithm requires only binary information indicating whether a user was satisfied with each of the recommended item or not. This makes the algorithm widely appealing to real world issues such as identification of search engine bias and pharmaceutical lobbying. We prove that the proposed algorithm detects the bias with high probability for a broad class of recommendation systems when sufficient number of users provide feedback on sufficient number of recommendations. We provide extensive simulations with real data sets and practical recommender systems, which confirm the trade offs in the theoretical guarantees.
 
Article
We consider streaming over a peer-to-peer network with homogeneous nodes in which a single source broadcasts a data stream to all the users in the system. Peers are allowed to enter or leave the system (adversarially) arbitrarily. Previous approaches for streaming in this setting have either used randomized distribution graphs or structured trees with randomized maintenance algorithms. Randomized graphs handle peer churn well but have poor connectivity guarantees, while structured trees have good connectivity but have proven hard to maintain under peer churn. We improve upon both approaches by presenting a novel distribution structure with a deterministic and distributed algorithm for maintenance under peer churn; our result is inspired by a recent work proposing deterministic algorithms for rumor spreading in graphs. A key innovation in our approach is in having redundant links in the distribution structure. While this leads to a reduction in the maximum streaming rate possible, we show that for the amount of redundancy used, the delay guarantee of the proposed algorithm is near optimal. We introduce a tolerance parameter that captures the worst-case transient streaming rate received by the peers during churn events and characterize the fundamental tradeoff between rate, delay and tolerance. A natural generalization of the deterministic algorithm achieves this tradeoff near optimally. Finally, the proposed deterministic algorithm is robust enough to handle various generalizations: ability to deal with heterogeneous node capacities of the peers and more complicated streaming patterns where multiple source transmissions are present.
 
Article
With traditional event list techniques, evaluating a detailed discrete event simulation model can often require hours or even days of computation time. Parallel simulation mimics the interacting servers and queues of a real system by assigning each simulated entity to a processor. By eliminating the event list and maintaining only sufficient synchronization to insure causality, parallel simulation can potentially provide speedups that are linear in the number of processors. A set of shared memory experiments is presented using the Chandy-Misra distributed simulation algorithm to simulate networks of queues. Parameters include queueing network topology and routing probabilities, number of processors, and assignment of network nodes to processors. These experiments show that Chandy-Misra distributed simulation is a questionable alternative to sequential simulation of most queueing network models.
 
Example of three synchronized nodes using the Trickle algorithm. 
One node carries the complete transmission load of the network. 
Article
As the use of wireless sensor networks increases, the need for (energy-)efficient and reliable broadcasting algorithms grows. Ideally, a broadcasting algorithm should have the ability to quickly disseminate data, while keeping the number of transmissions low. In this paper we develop a model describing the message count in large-scale wireless sensor networks. We focus our attention on the popular Trickle algorithm, which has been proposed as a suitable communication protocol for code maintenance and propagation in wireless sensor networks. Besides providing a mathematical analysis of the algorithm, we propose a generalized version of Trickle, with an additional parameter defining the length of a listen-only period. This generalization proves to be useful for optimizing the design and usage of the algorithm. For single-cell networks we show how the message count increases with the size of the network and how this depends on the Trickle parameters. Furthermore, we derive distributions of inter-broadcasting times and investigate their asymptotic behavior. Our results prove conjectures made in the literature concerning the effect of a listen-only period. Additionally, we develop an approximation for the expected number of transmissions in multi-cell networks. All results are validated by simulations.
 
Article
Recent advances have resulted in queue-based algorithms for medium access control which operate in a distributed fashion, and yet achieve the optimal throughput performance of centralized scheduling algorithms. However, fundamental performance bounds reveal that the "cautious" activation rules involved in establishing throughput optimality tend to produce extremely large delays, typically growing exponentially in 1/(1-r), with r the load of the system, in contrast to the usual linear growth. Motivated by that issue, we explore to what extent more "aggressive" schemes can improve the delay performance. Our main finding is that aggressive activation rules induce a lingering effect, where individual nodes retain possession of a shared resource for excessive lengths of time even while a majority of other nodes idle. Using central limit theorem type arguments, we prove that the idleness induced by the lingering effect may cause the delays to grow with 1/(1-r) at a quadratic rate. To the best of our knowledge, these are the first mathematical results illuminating the lingering effect and quantifying the performance impact. In addition extensive simulation experiments are conducted to illustrate and validate the various analytical results.
 
Article
Among the many techniques in computer graphics, ray tracing is prized because it can render realistic images, albeit at great computational expense. In this note, the performance of several approaches to ray tracing on a distributed memory parallel system is evaluated. A set of performance instrumentation tools and their associated visualization software are used to identify the underlying causes of performance differences.
 
Article
Wireless network topologies change over time and maintaining routes requires frequent updates. Updates are costly in terms of consuming throughput available for data transmission, which is precious in wireless networks. In this paper, we ask whether there exist low-overhead schemes that produce low-stretch routes. This is studied by using the underlying geometric properties of the connectivity graph in wireless networks. Comment: 29 pages, 19 figures, a shorter version was published in the proceedings of the 2008 ACM Sigmetrics conference
 
Number of events generated by CFTP algortihms for the star graph with 100 vertices for different values of λ.  
Number of events generated by CFTP algortihms for the star graph with 1000 vertices for different values of λ.  
Number of events generated by CFTP algortihms for the Barabasi-Albert model with 100 vertices for different values of λ.  
Article
The maximum independent set (MIS) problem is a well-studied combinatorial optimization problem that naturally arises in many applications, such as wireless communication, information theory and statistical mechanics. MIS problem is NP-hard, thus many results in the literature focus on fast generation of maximal independent sets of high cardinality. One possibility is to combine Gibbs sampling with coupling from the past arguments to detect convergence to the stationary regime. This results in a sampling procedure with time complexity that depends on the mixing time of the Glauber dynamics Markov chain. We propose an adaptive method for random event generation in the Glauber dynamics that considers only the events that are effective in the coupling from the past scheme, accelerating the convergence time of the Gibbs sampling algorithm. The full paper is available on arXiv.
 
Article
Network service providers and customers are often concerned with aggregate performance measures that span multiple network paths. Unfortunately, forming such network-wide measures can be difficult, due to the issues of scale involved. In particular, the number of paths grows too rapidly with the number of endpoints to make exhaustive measurement practical. As a result, there is interest in the feasibility of methods that dramatically reduce the number of paths measured in such situations while maintaining acceptable accuracy. In previous work we proposed a statistical framework to efficiently address this problem, in the context of additive metrics such as delay and loss rate, for which the per-path metric is a sum of (possibly transformed) per-link measures. The key to our method lies in the observation and exploitation of significant redundancy in network paths (sharing of common links). In this paper we make three contributions: (1) we generalize the framework to make it more immediately applicable to network measurements encountered in practice; (2) we demonstrate that the observed path redundancy upon which our method is based is robust to variation in key network conditions and characteristics, including link failures; and (3) we show how the framework may be applied to address three practical problems of interest to network providers and customers, using data from an operating network. In particular, we show how appropriate selection of small sets of path measurements can be used to accurately estimate network-wide averages of path delays, to reliably detect network anomalies, and to effectively make a choice between alternative sub-networks, as a customer choosing between two providers or two ingress points into a provider network.
 
Top-cited authors
Kai Li
  • Harbin Institute of Technology
Adam Wierman
  • California Institute of Technology
Giuliana Franceschinis
  • Amedeo Avogadro University of Eastern Piedmont
Gianfranco Balbo
  • Università degli Studi di Torino
Gianni Conte
  • Università di Parma