Skip to main content

Xiaoxi Zhang

Followers

2

Following

1

Public Views

Negin Golrezaei

Ilqar Ramazanli

Carnegie Mellon University

Zinovi Rabinovich

Claudio Gambella

Università di Bologna

Aditya Regiment

KMUTNB

Panagiotis Andrianesis

Technical University of Munich

Interests

Uploads

Papers by Xiaoxi Zhang

Machine Learning on Volatile Instances

IEEE INFOCOM 2020 - IEEE Conference on Computer Communications, 2020

Due to the massive size of the neural network models and training datasets used in machine learni... more Due to the massive size of the neural network models and training datasets used in machine learning today, it is imperative to distribute stochastic gradient descent (SGD) by splitting up tasks such as gradient evaluation across multiple worker nodes. However, running distributed SGD can be prohibitively expensive because it may require specialized computing resources such as GPUs for extended periods of time. We propose cost-effective strategies to exploit volatile cloud instances that are cheaper than standard instances, but may be interrupted by higher priority workloads. To the best of our knowledge, this work is the first to quantify how variations in the number of active worker nodes (as a result of preemption) affects SGD convergence and the time to train the model. By understanding these tradeoffs between preemption probability of the instances, accuracy, and training time, we are able to derive practical strategies for configuring distributed SGD jobs on volatile instances such as Amazon EC2 spot instances and other preemptible cloud instances. Experimental results show that our strategies achieve good training performance at substantially lower cost.

CASTLE over the Air

Proceedings of the 17th Annual International Conference on Mobile Systems, Applications, and Services, 2019

This paper presents a fully distributed scheduling framework called CASTLE (Client-side Adaptive ... more This paper presents a fully distributed scheduling framework called CASTLE (Client-side Adaptive Scheduler That minimizes Load and Energy), which jointly optimizes the spectral efficiency of cellular networks and battery consumption of smart devices. To do so, we focus on scenarios when many smart devices compete for cellular resources in the same base station: spreading out transmissions over time so that only a few devices transmit at once improves both spectral efficiency and battery consumption. To this end, we devise two novel features in CASTLE. First, we explicitly consider intercell interference for accurate cellular load estimation. Based on our observations, we exploit the RSRQ (Reference Signal Received Quality) and SINR as features in a machine learning algorithm to accurately estimate the cellular load. Second, we propose a fully distributed scheduling algorithm that coordinates transmissions between clients based on the locally estimated load level at each client. Our formulation for minimizing battery consumption at each device leads to an optimized backoff-based algorithm that fits practical environments. To evaluate these features, we prototype a complete LTE system testbed consisting of mobile devices, eNodeBs, EPC (Evolved Packet Core) and application servers. Our comprehensive experimental results show that CASTLE's load estimation is up to 91% accurate, and that CASTLE achieves higher spectral efficiency with less battery consumption, compared to existing centralized

Optimal Learning-Based Network Protocol Selection

Today’s Internet must support applications with increasingly dynamic and heterogeneous connectivi... more Today’s Internet must support applications with increasingly dynamic and heterogeneous connectivity requirements, such as video streaming, virtual reality (VR), and the Internet of Things. Yet current network management practices generally rely on pre-specified flow configurations, which may not be able to cope with dynamic flow priorities or changing network conditions, e.g., on volatile wireless links. In this work, we instead propose a model-free learning approach to find the optimal network policies for current flow requirements. This approach is attractive as comprehensive models do not exist for how different protocol choices affect flow performance, which may further be affected by dynamically changing network conditions. However, it raises new technical challenges: policy configurations can affect the performance of multiple flows sharing the network resources, and this flow coupling cannot be readily handled by existing online learning algorithms. In this work, we extend mu...

MoDEMS: Optimizing Edge Computing Migrations For User Mobility

2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS), 2021

Edge computing capabilities in 5G wireless networks promise to benefit mobile users: computing ta... more Edge computing capabilities in 5G wireless networks promise to benefit mobile users: computing tasks can be offloaded from user devices to nearby edge servers, reducing users' experienced latencies. Few works have addressed how this offloading should handle long-term user mobility: as devices move, they will need to offload to different edge servers, which may require migrating data or state information from one edge server to another. In this paper, we introduce MoDEMS, a system model and architecture that provides a rigorous theoretical framework and studies the challenges of such migrations to minimize the service provider cost and user latency. We show that this cost minimization problem can be expressed as an integer linear programming problem, which is hard to solve due to resource constraints at the servers and unknown user mobility patterns. We show that finding the optimal migration plan is in general NPhard, and we propose alternative heuristic solution algorithms that perform well in both theory and practice. We finally validate our results with real user mobility traces, ns-3 simulations, and an LTE testbed experiment. Migrations reduce the latency experienced by users of edge applications by 33% compared to previously proposed migration approaches.

Online Stochastic Buy-Sell Mechanism for VNF Chains in the NFV Market

IEEE Journal on Selected Areas in Communications, 2017

With the recent advent of network functions virtualization (NFV), enterprises and businesses are ... more With the recent advent of network functions virtualization (NFV), enterprises and businesses are looking into network service provisioning through the service chains of virtual network functions (VNFs), instead of relying on dedicated hardware middleboxes. Accompanying this trend, an NFV market is emerging, where NFV service providers create VNF instances, assemble VNF service chains, and sell them for the use of customers, using resources (computing, bandwidth) that they own or rent from other resource suppliers. Efficient service chain provisioning and pricing mechanisms are still missing, to charge assembled service chains according to demand and the supply of resources at any time. We propose an online stochastic auction mechanism for on-demand service chain provisioning and pricing at an NFV provider. Our auction takes in buy bids for service chains from multiple customers and sell bids from various resource suppliers to supplement the NFV provider's geodistributed resource pool, with resource occupation/contribution durations. We extend online primal-dual optimization framework for handling both buyers and sellers, with a new competitive analysis. The online mechanism maximizes the expected social welfare of the NFV ecosystem (the NFV provider, customers and resource suppliers) with a good competitive ratio as compared with the expected offline optimal social welfare, while guaranteeing truthfulness in bidding, individual rationality for both buyers and sellers, and polynomial time for computation. We evaluate our mechanism through trace-driven simulation studies, and demonstrate a close-to-offline-optimal performance in expected social welfare under realistic settings.

Online Auctions in IaaS Clouds

ACM SIGMETRICS Performance Evaluation Review, 2015

Auction design has recently been studied for dynamic resource bundling and VM provisioning in Iaa... more Auction design has recently been studied for dynamic resource bundling and VM provisioning in IaaS clouds, but is mostly restricted to the one-shot or offline setting. This work targets a more realistic case of online VM auction design, where: (i) cloud users bid for resources into the future to assemble customized VMs with desired occupation durations; (ii) the cloud provider dynamically packs multiple types of resources on heterogeneous physical machines (servers) into the requested VMs; (iii) the operational costs of servers are considered in resource allocation; (iv) both social welfare and the cloud provider's net profit are to be maximized over the system running span. We design truthful, polynomial time auctions to achieve social welfare maximization and/or the provider's profit maximization with good competitive ratios. Our mechanisms consist of two main modules: (1) an online primal-dual optimization framework for VM allocation to maximize the social welfare with se...

On Space Information Flow: Single multicast

2013 International Symposium on Network Coding (NetCod), 2013

Departing from Network Information Flow (NIF) that studies network coding in graphs, Space Inform... more Departing from Network Information Flow (NIF) that studies network coding in graphs, Space Information Flow (SIF) is a new paradigm that studies network coding in a geometric space. This work focuses on the problem of min-cost multicast network coding in a 2-dimensional Euclidean space. We prove a number of properties of the optimal SIF solutions, and propose a two-phase heuristic algorithm for computing the optimal SIF. The first phase computes the optimal topology through space partitioning that translates the SIF problem into a NIF problem, which is then solved using linear optimization. The second phase computes the min-cost embedding of the SIF topology found in the first phase, by fine tuning the location of each relay node using properties that an optimal SIF must satisfy. I. INTRODUCTION Network Information Flow (NIF) [1], which studies network coding in graphs, was proposed in 2000. The main underlying technique of NIF is network coding, which can improve the throughput of a network and reduce the complexity of computing the optimal transmission scheme [2]. The ratio of the maximum throughput achievable with network coding over that with routing is known as the coding advantage [3].

Occupation-Oblivious Pricing of Cloud Jobs via Online Learning

IEEE INFOCOM 2018 - IEEE Conference on Computer Communications, 2018

State-of-the-art cloud platforms adopt pay-as-you-go pricing, where users pay for the resources o... more State-of-the-art cloud platforms adopt pay-as-you-go pricing, where users pay for the resources on demand according to occupation time. Simple and intuitive as it is, such a pricing scheme is a mismatch for new workloads today such as largescale machine learning, whose completion time is hard to estimate beforehand. To supplement existing cloud pricing schemes, we propose an occupation-oblivious online pricing mechanism for cloud jobs without pre-specified time duration and for users who prefer a predetermined cost for job execution. Our strategy posts unit resource prices upon user arrival and decides a fixed charge for completing the user's job, without the need to know how long the job is to occupy the requested resources. At the core of our design is a novel multi-armed bandit based online learning algorithm for estimating unknown input by exploration and exploitation of past resource sales, and deciding resource prices to maximize profit of the cloud provider in an online setting. Our online learning algorithm achieves a low regret sublinear with the time horizon, in terms of overall provider profit, compared with an omniscient benchmark. We also conduct tracedriven simulations to verify efficacy of the algorithm in real-world settings.

Online cost minimization for operating geo-distributed cloud CDNs

2015 IEEE 23rd International Symposium on Quality of Service (IWQoS), 2015

Cloud-based content delivery networks (Cloud CDN) cache and deliver contents from geo-distributed... more Cloud-based content delivery networks (Cloud CDN) cache and deliver contents from geo-distributed cloud data centers to end users across the globe, exploiting "infinite" ondemand cloud resources to address volatile user demands. It is critically important to efficiently manage cloud resources in different locations over time, for minimization of the operational cost of the CDN provider, while delivering short response delay to user requests. Although many have studied cost-aware replica placement and request redirection in CDN systems, most are restricted to an offline or one-time setting, or resort to greedy heuristics for online operation. This work proposes an efficient online algorithm for dynamic content replication and request dispatching in cloud CDNs operating over a long time span, targeting overall cost minimization with performance guarantees. Our online algorithm consists of two main modules: (1) a regularization method from the online learning literature to convert the offline cost-minimization optimization problem into a sequence of regularized problems, each to be efficiently solvable in one time slot; (2) a randomized approach to convert the optimal fractional solutions from the regularized problems to integer solutions of the original problem, achieving a good competitive ratio. The effectiveness of our online algorithm is validated through solid theoretical analysis and trace-driven simulations.

Machine Learning on Volatile Instances

IEEE INFOCOM 2020 - IEEE Conference on Computer Communications, 2020

Due to the massive size of the neural network models and training datasets used in machine learni... more Due to the massive size of the neural network models and training datasets used in machine learning today, it is imperative to distribute stochastic gradient descent (SGD) by splitting up tasks such as gradient evaluation across multiple worker nodes. However, running distributed SGD can be prohibitively expensive because it may require specialized computing resources such as GPUs for extended periods of time. We propose cost-effective strategies to exploit volatile cloud instances that are cheaper than standard instances, but may be interrupted by higher priority workloads. To the best of our knowledge, this work is the first to quantify how variations in the number of active worker nodes (as a result of preemption) affects SGD convergence and the time to train the model. By understanding these tradeoffs between preemption probability of the instances, accuracy, and training time, we are able to derive practical strategies for configuring distributed SGD jobs on volatile instances such as Amazon EC2 spot instances and other preemptible cloud instances. Experimental results show that our strategies achieve good training performance at substantially lower cost.

CASTLE over the Air

Proceedings of the 17th Annual International Conference on Mobile Systems, Applications, and Services, 2019

This paper presents a fully distributed scheduling framework called CASTLE (Client-side Adaptive ... more This paper presents a fully distributed scheduling framework called CASTLE (Client-side Adaptive Scheduler That minimizes Load and Energy), which jointly optimizes the spectral efficiency of cellular networks and battery consumption of smart devices. To do so, we focus on scenarios when many smart devices compete for cellular resources in the same base station: spreading out transmissions over time so that only a few devices transmit at once improves both spectral efficiency and battery consumption. To this end, we devise two novel features in CASTLE. First, we explicitly consider intercell interference for accurate cellular load estimation. Based on our observations, we exploit the RSRQ (Reference Signal Received Quality) and SINR as features in a machine learning algorithm to accurately estimate the cellular load. Second, we propose a fully distributed scheduling algorithm that coordinates transmissions between clients based on the locally estimated load level at each client. Our formulation for minimizing battery consumption at each device leads to an optimized backoff-based algorithm that fits practical environments. To evaluate these features, we prototype a complete LTE system testbed consisting of mobile devices, eNodeBs, EPC (Evolved Packet Core) and application servers. Our comprehensive experimental results show that CASTLE's load estimation is up to 91% accurate, and that CASTLE achieves higher spectral efficiency with less battery consumption, compared to existing centralized

Optimal Learning-Based Network Protocol Selection

Today’s Internet must support applications with increasingly dynamic and heterogeneous connectivi... more Today’s Internet must support applications with increasingly dynamic and heterogeneous connectivity requirements, such as video streaming, virtual reality (VR), and the Internet of Things. Yet current network management practices generally rely on pre-specified flow configurations, which may not be able to cope with dynamic flow priorities or changing network conditions, e.g., on volatile wireless links. In this work, we instead propose a model-free learning approach to find the optimal network policies for current flow requirements. This approach is attractive as comprehensive models do not exist for how different protocol choices affect flow performance, which may further be affected by dynamically changing network conditions. However, it raises new technical challenges: policy configurations can affect the performance of multiple flows sharing the network resources, and this flow coupling cannot be readily handled by existing online learning algorithms. In this work, we extend mu...

MoDEMS: Optimizing Edge Computing Migrations For User Mobility

2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS), 2021

Edge computing capabilities in 5G wireless networks promise to benefit mobile users: computing ta... more Edge computing capabilities in 5G wireless networks promise to benefit mobile users: computing tasks can be offloaded from user devices to nearby edge servers, reducing users' experienced latencies. Few works have addressed how this offloading should handle long-term user mobility: as devices move, they will need to offload to different edge servers, which may require migrating data or state information from one edge server to another. In this paper, we introduce MoDEMS, a system model and architecture that provides a rigorous theoretical framework and studies the challenges of such migrations to minimize the service provider cost and user latency. We show that this cost minimization problem can be expressed as an integer linear programming problem, which is hard to solve due to resource constraints at the servers and unknown user mobility patterns. We show that finding the optimal migration plan is in general NPhard, and we propose alternative heuristic solution algorithms that perform well in both theory and practice. We finally validate our results with real user mobility traces, ns-3 simulations, and an LTE testbed experiment. Migrations reduce the latency experienced by users of edge applications by 33% compared to previously proposed migration approaches.

Online Stochastic Buy-Sell Mechanism for VNF Chains in the NFV Market

IEEE Journal on Selected Areas in Communications, 2017

With the recent advent of network functions virtualization (NFV), enterprises and businesses are ... more With the recent advent of network functions virtualization (NFV), enterprises and businesses are looking into network service provisioning through the service chains of virtual network functions (VNFs), instead of relying on dedicated hardware middleboxes. Accompanying this trend, an NFV market is emerging, where NFV service providers create VNF instances, assemble VNF service chains, and sell them for the use of customers, using resources (computing, bandwidth) that they own or rent from other resource suppliers. Efficient service chain provisioning and pricing mechanisms are still missing, to charge assembled service chains according to demand and the supply of resources at any time. We propose an online stochastic auction mechanism for on-demand service chain provisioning and pricing at an NFV provider. Our auction takes in buy bids for service chains from multiple customers and sell bids from various resource suppliers to supplement the NFV provider's geodistributed resource pool, with resource occupation/contribution durations. We extend online primal-dual optimization framework for handling both buyers and sellers, with a new competitive analysis. The online mechanism maximizes the expected social welfare of the NFV ecosystem (the NFV provider, customers and resource suppliers) with a good competitive ratio as compared with the expected offline optimal social welfare, while guaranteeing truthfulness in bidding, individual rationality for both buyers and sellers, and polynomial time for computation. We evaluate our mechanism through trace-driven simulation studies, and demonstrate a close-to-offline-optimal performance in expected social welfare under realistic settings.

Online Auctions in IaaS Clouds

ACM SIGMETRICS Performance Evaluation Review, 2015

Auction design has recently been studied for dynamic resource bundling and VM provisioning in Iaa... more Auction design has recently been studied for dynamic resource bundling and VM provisioning in IaaS clouds, but is mostly restricted to the one-shot or offline setting. This work targets a more realistic case of online VM auction design, where: (i) cloud users bid for resources into the future to assemble customized VMs with desired occupation durations; (ii) the cloud provider dynamically packs multiple types of resources on heterogeneous physical machines (servers) into the requested VMs; (iii) the operational costs of servers are considered in resource allocation; (iv) both social welfare and the cloud provider's net profit are to be maximized over the system running span. We design truthful, polynomial time auctions to achieve social welfare maximization and/or the provider's profit maximization with good competitive ratios. Our mechanisms consist of two main modules: (1) an online primal-dual optimization framework for VM allocation to maximize the social welfare with se...

On Space Information Flow: Single multicast

2013 International Symposium on Network Coding (NetCod), 2013

Departing from Network Information Flow (NIF) that studies network coding in graphs, Space Inform... more Departing from Network Information Flow (NIF) that studies network coding in graphs, Space Information Flow (SIF) is a new paradigm that studies network coding in a geometric space. This work focuses on the problem of min-cost multicast network coding in a 2-dimensional Euclidean space. We prove a number of properties of the optimal SIF solutions, and propose a two-phase heuristic algorithm for computing the optimal SIF. The first phase computes the optimal topology through space partitioning that translates the SIF problem into a NIF problem, which is then solved using linear optimization. The second phase computes the min-cost embedding of the SIF topology found in the first phase, by fine tuning the location of each relay node using properties that an optimal SIF must satisfy. I. INTRODUCTION Network Information Flow (NIF) [1], which studies network coding in graphs, was proposed in 2000. The main underlying technique of NIF is network coding, which can improve the throughput of a network and reduce the complexity of computing the optimal transmission scheme [2]. The ratio of the maximum throughput achievable with network coding over that with routing is known as the coding advantage [3].

Occupation-Oblivious Pricing of Cloud Jobs via Online Learning

IEEE INFOCOM 2018 - IEEE Conference on Computer Communications, 2018

State-of-the-art cloud platforms adopt pay-as-you-go pricing, where users pay for the resources o... more State-of-the-art cloud platforms adopt pay-as-you-go pricing, where users pay for the resources on demand according to occupation time. Simple and intuitive as it is, such a pricing scheme is a mismatch for new workloads today such as largescale machine learning, whose completion time is hard to estimate beforehand. To supplement existing cloud pricing schemes, we propose an occupation-oblivious online pricing mechanism for cloud jobs without pre-specified time duration and for users who prefer a predetermined cost for job execution. Our strategy posts unit resource prices upon user arrival and decides a fixed charge for completing the user's job, without the need to know how long the job is to occupy the requested resources. At the core of our design is a novel multi-armed bandit based online learning algorithm for estimating unknown input by exploration and exploitation of past resource sales, and deciding resource prices to maximize profit of the cloud provider in an online setting. Our online learning algorithm achieves a low regret sublinear with the time horizon, in terms of overall provider profit, compared with an omniscient benchmark. We also conduct tracedriven simulations to verify efficacy of the algorithm in real-world settings.

Online cost minimization for operating geo-distributed cloud CDNs

2015 IEEE 23rd International Symposium on Quality of Service (IWQoS), 2015

Cloud-based content delivery networks (Cloud CDN) cache and deliver contents from geo-distributed... more Cloud-based content delivery networks (Cloud CDN) cache and deliver contents from geo-distributed cloud data centers to end users across the globe, exploiting "infinite" ondemand cloud resources to address volatile user demands. It is critically important to efficiently manage cloud resources in different locations over time, for minimization of the operational cost of the CDN provider, while delivering short response delay to user requests. Although many have studied cost-aware replica placement and request redirection in CDN systems, most are restricted to an offline or one-time setting, or resort to greedy heuristics for online operation. This work proposes an efficient online algorithm for dynamic content replication and request dispatching in cloud CDNs operating over a long time span, targeting overall cost minimization with performance guarantees. Our online algorithm consists of two main modules: (1) a regularization method from the online learning literature to convert the offline cost-minimization optimization problem into a sequence of regularized problems, each to be efficiently solvable in one time slot; (2) a randomized approach to convert the optimal fractional solutions from the regularized problems to integer solutions of the original problem, achieving a good competitive ratio. The effectiveness of our online algorithm is validated through solid theoretical analysis and trace-driven simulations.