arXiv subjects

Emina Soljanin

Publications and source records attributed to Emina Soljanin.

At least 55 records · Page 3Linked to original sources

arXiv2019

Scheduling in the Presence of Data Intensive Compute Jobs

We study the performance of non-adaptive scheduling policies in computing systems with multiple servers. Compute jobs are mostly regular, with modest service requirements. However, there are sporadic data intensive jobs, whose expected service time is much higher than that of the regular jobs. Forthis model, we are interested in the effect of scheduling policieson the average time a job spends in the system. To this end, we introduce two performance indicators in a simplified, only-arrival system. We believe that these performance indicators are good predictors of the relative performance of the policies in the queuing system, which is supported by simulations results.

cs.PF↗

arXiv2019

Data Replication for Reducing Computing Time in Distributed Systems with Stragglers

In distributed computing systems with stragglers, various forms of redundancy can improve the average delay performance. We study the optimal replication of data in systems where the job execution time is a stochastically decreasing and convex random variable. We show that in such systems, the optimum assignment policy is the balanced replication of disjoint batches of data. Furthermore, for Exponential and Shifted-Exponential service times, we derive the optimum redundancy levels for minimizing both expected value and the variance of the job completion time. Our analysis shows that, the optimum redundancy level may not be the same for the two metrics, thus there is a trade-off between reducing the expected value of the completion time and reducing its variance.

cs.DC↗

arXiv2019

Straggler Mitigation at Scale

Runtime performance variability at the servers has been a major issue, hindering the predictable and scalable performance in modern distributed systems. Executing requests or jobs redundantly over multiple servers has been shown to be effective for mitigating variability, both in theory and practice. Systems that employ redundancy has drawn significant attention, and numerous papers have analyzed the pain and gain of redundancy under various service models and assumptions on the runtime variability. This paper presents a cost (pain) vs. latency (gain) analysis of executing jobs of many tasks by employing replicated or erasure coded redundancy. Tail heaviness of service time variability is decisive on the pain and gain of redundancy and we quantify its effect by deriving expressions for the cost and latency. Specifically, we try to answer four questions: 1) How do replicated and coded redundancy compare in the cost vs. latency tradeoff? 2) Can we introduce redundancy after waiting some time and expect to reduce the cost? 3) Can relaunching the tasks that appear to be straggling after some time help to reduce cost and/or latency? 4) Is it effective to use redundancy and relaunching together? We validate the answers we found for each of the questions via simulations that use empirical distributions extracted from a Google cluster data.

cs.PF↗

arXiv2019

Redundancy Scheduling in Systems with Bi-Modal Job Service Time Distribution

Queuing systems with redundant requests have drawn great attention because of their promise to reduce the job completion time and variability. Despite a large body of work on the topic, we are still far from fully understanding the benefits of redundancy in practice. We here take one step towards practical systems by studying queuing systems with bi-modal job service time distribution. Such distributions have been observed in practice, as can be seen in, e.g., Google cluster traces. We develop an analogy to a classical urns and balls problem, and use it to study the queuing time performance of two non-adaptive classical scheduling policies: random and round-robin. We introduce new performance indicators in the analogous model, and argue that they are good predictors of the queuing time in non-adaptive scheduling policies. We then propose a non-adaptive scheduling policy that is based on combinatorial designs, and show that it has better performance indicators. Simulations confirm that the proposed scheduling policy, as the performance indicators suggest, reduces the queuing times compared to random and round-robin scheduling.

cs.PF↗

arXiv2019

Straggling for Covert Message Passing on Complete Graphs

We introduce a model for mobile, multi-agent information transfer that increases the communication covertness through a protocol which also increases the information transfer delay. Covertness is achieved in the presence of a warden who has the ability to patrol the communication channels. Furthermore we show how two forms of redundancy can be used as an effective tool to control the tradeoff between the covertness and the delay.

cs.CR↗

arXiv2019

Anonymity Mixes as (Partial) Assembly Queues: Modeling and Analysis

Anonymity platforms route the traffic over a network of special routers that are known as mixes and implement various traffic disruption techniques to hide the communicating users' identities. Batch mixes in particular anonymize communicating peers by allowing message exchange to take place only after a sufficient number of messages (a batch) accumulate, thus introducing delay. We introduce a queueing model for batch mix and study its delay properties. Our analysis shows that delay of a batch mix grows quickly as the batch size gets close to the number of senders connected to the mix. We then propose a randomized batch mixing strategy and show that it achieves much better delay scaling in terms of the batch size. However, randomization is shown to reduce the anonymity preserving capabilities of the mix. We also observe that queueing models are particularly useful to study anonymity metrics that are more practically relevant such as the time-to-deanonymize metric.

cs.PF↗

arXiv2019

Optimizing Redundancy Levels in Master-Worker Compute Clusters for Straggler Mitigation

Runtime variability in computing systems causes some tasks to straggle and take much longer than expected to complete. These straggler tasks are known to significantly slowdown distributed computation. Job execution with speculative execution of redundant tasks has been the most widely deployed technique for mitigating the impact of stragglers, and many recent theoretical papers have studied the advantages and disadvantages of using redundancy under various system and service models. However, no clear guidelines could yet be found on when, for which jobs, and how much redundancy should be employed in Master-Worker compute clusters, which is the most widely adopted architecture in modern compute systems. We are concerned with finding a strategy for scheduling jobs with redundancy that works well in practice. This is a complex optimization problem, which we address in stages. We first use Reinforcement Learning (RL) techniques to learn good scheduling principles from realistic experience. Building on these principles, we derive a simple scheduling policy and present an approximate analysis of its performance. Specifically, we derive expressions to decide when and which jobs should be scheduled with how much redundancy. We show that policy that we devise in this way performs as good as the more complex policies that are derived by RL. Finally, we extend our approximate analysis to the case when system employs the other widely deployed remedy for stragglers, which is relaunching straggler tasks after waiting some time. We show that scheduling with redundancy significantly outperforms straggler relaunch policy when the offered load on the system is low or moderate, and performs slightly worse when the offered load is very high.

cs.PF↗

arXiv2019

Service Rate Region of Content Access from Erasure Coded Storage

We consider storage systems in which $K$ files are stored over $N$ nodes. A node may be systematic for a particular file in the sense that access to it gives access to the file. Alternatively, a node may be coded, meaning that it gives access to a particular file only when combined with other nodes (which may be coded or systematic). Requests for file $f_k$ arrive at rate $λ_k$, and we are interested in the rate that can be served by a particular system. In this paper, we determine the set of request arrival rates for the a $3$-file coded storage system. We also provide an algorithm to maximize the rate of requests served for file $K$ given $λ_1,\dots, λ_{K-1}$ in a general $K$-file case.

cs.IT↗

arXiv2018

Timely Updates over an Erasure Channel

Using an age of information (AoI) metric, we examine the transmission of coded updates through a binary erasure channel to a monitor/receiver. We start by deriving the average status update age of an infinite incremental redundancy (IIR) system in which the transmission of a k-symbol update continuesuntil k symbols are received. This system is then compared to a fixed redundancy (FR) system in which each update is transmitted as an n symbol packet and the packet is successfully received if and only if at least k symbols are received. If fewer than k symbols are received, the update is discarded. Unlike the IIR system, the FR system requires no feedback from the receiver. For a single monitor system, we show that tuning the redundancy to the symbol erasure rate enables the FR system to perform as well as the IIR system. As the number of monitors is increased, the FR system outperforms the IIR system that guarantees delivery of all updates to all monitors.

cs.IT↗

arXiv2018

Timely Lossless Source Coding for Randomly Arriving Symbols

We consider a real-time streaming source coding system in which an encoder observes a sequence of randomly arriving symbols from an i.i.d. source, and feeds binary codewords to a FIFO buffer that outputs one bit per time unit to a decoder. Each source symbol represents a status update by the source, and the timeliness of the system is quantified by the age of information (AoI), defined as the time difference between the present time and the generation time of the most up-to-date symbol at the output of the decoder. When the FIFO buffer is allowed to be empty, we propose an optimal prefix-free lossless coding scheme that minimizes the average peak age based on the analysis of discrete-time Geo/G/1 queue. For more practical scenarios in which a special codeword is reserved for indicating an empty buffer, we propose an encoding scheme that assigns a codeword to the empty buffer state based on an estimate of the buffer idle time.

cs.IT↗

arXiv2018

Status Updates Through Multicast Networks

Using age of information as the freshness metric, we examine a multicast network in which real-time status updates are generated by the source and sent to a group of $n$ interested receivers. We show that in order to keep the information freshness at each receiver, the source should terminate the transmission of the current update and start sending a new update packet as soon as it receives the acknowledgements back from any $k$ out of $n$ nodes. As the source stopping threshold $k$ increases, a node is more likely to get the latest generated update, but the age of the most recent update is more likely to become outdated. We derive the age minimized stopping threshold $k$ that balances the likelihood of getting the latest update and the freshness of the latest update for shifted exponential link delay. Through numerical evaluations for different stopping strategies, we find that waiting for the acknowledgements from the earliest $k$ out of $n$ nodes leads to lower average age than waiting for a pre-selected group of $k$ nodes. We also observe that a properly chosen threshold $k$ can prevent information staleness for increasing number of nodes $n$ in the multicast network.

cs.IT↗

arXiv2018

Multicast With Prioritized Delivery: How Fresh is Your Data?

We consider a multicast network in which real-time status updates generated by a source are replicated and sent to multiple interested receiving nodes through independent links. The receiving nodes are divided into two groups: one priority group consists of $k$ nodes that require the reception of every update packet, the other non-priority group consists of all other nodes without the delivery requirement. Using age of information as a freshness metric, we analyze the time-averaged age at both priority and non-priority nodes. For shifted-exponential link delay distributions, the average age at a priority node is lower than that at a non-priority node due to the delivery guarantee. However, this advantage for priority nodes disappears if the link delay is exponential distributed. Both groups of nodes have the same time-averaged age, which implies that the guaranteed delivery of updates has no effect the time-averaged freshness.

cs.IT↗

arXiv2018

On the Effect of Task-to-Worker Assignment in Distributed Computing Systems with Stragglers

We study the expected completion time of some recently proposed algorithms for distributed computing which redundantly assign computing tasks to multiple machines in order to tolerate a certain number of machine failures. We analytically show that not only the amount of redundancy but also the task-to-machine assignments affect the latency in a distributed system. We study systems with a fixed number of computing tasks that are split in possibly overlapping batches, and independent exponentially distributed machine service times. We show that, for such systems, the uniform replication of non- overlapping (disjoint) batches of computing tasks achieves the minimum expected computing time.

cs.DC↗

arXiv2018

On Distributed Storage Allocations of Large Files for Maximum Service Rate

Allocation of (redundant) file chunks throughout a distributed storage system affects important performance metrics such as the probability of file recovery, data download time, or the service rate of the system under a given data access model. This paper is concerned with the service rate under the assumption that the stored data is large and its download time is not negligible. We focus on quasi-uniform storage allocations and provide a service rate analysis for two common data access models. We find that the optimal allocation varies in accordance with different system parameters. This was not the case under the assumption that the download time does not scale with the size of data, where the minimal spreading allocation was previously found to be universally optimal.

cs.DC↗

arXiv2018

Simplex Queues for Hot-Data Download

In cloud storage systems, hot data is usually replicated over multiple nodes in order to accommodate simultaneous access by multiple users as well as increase the fault tolerance of the system. Recent cloud storage research has proposed using availability codes, which is a special class of erasure codes, as a more storage efficient way to store hot data. These codes enable data recovery from multiple, small disjoint groups of servers. The number of the recovery groups is referred to as the availability and the size of each group as the locality of the code. Until now, we have very limited knowledge on how code locality and availability affect data access time. Data download from these systems involves multiple fork-join queues operating in-parallel, making the analysis of access time a very challenging problem. In this paper, we present an approximate analysis of data access time in storage systems that employ simplex codes, which are an important and in certain sense optimal class of availability codes. We consider and compare three strategies in assigning download requests to servers; first one aggressively exploits the storage availability for faster download, second one implements only load balancing, and the last one employs storage availability only for hot data download without incurring any negative impact on the cold data download.

cs.IT↗

arXiv2018

Minimizing Content Staleness in Dynamo-Style Replicated Storage Systems

Consistency in data storage systems requires any read operation to return the most recent written version of the content. In replicated storage systems, consistency comes at the price of delay due to large-scale write and read operations. Many applications with low latency requirements tolerate data staleness in order to provide high availability and low operation latency. Using age of information as the staleness metric, we examine a data updating system in which real-time content updates are replicated and stored in a Dynamo-style quorum- based distributed system. A source sends updates to all the nodes in the system and waits for acknowledgements from the earliest subset of nodes, known as a write quorum. An interested client fetches the update from another set of nodes, defined as a read quorum. We analyze the staleness-delay tradeoff in replicated storage by varying the write quorum size. With a larger write quorum, an instantaneous read is more likely to get the latest update written by the source. However, the age of the content written to the system is more likely to become stale as the write quorum size increases. For shifted exponential distributed write delay, we derive the age optimized write quorum size that balances the likelihood of reading the latest update and the freshness of the latest update written by the source.

cs.DC↗

arXiv2017

On the Service Capacity Region of Accessing Erasure Coded Content

Cloud storage systems generally add redundancy in storing content files such that $K$ files are replicated or erasure coded and stored on $N > K$ nodes. In addition to providing reliability against failures, the redundant copies can be used to serve a larger volume of content access requests. A request for one of the files can be either be sent to a systematic node, or one of the repair groups. In this paper, we seek to maximize the service capacity region, that is, the set of request arrival rates for the $K$ files that can be supported by a coded storage system. We explore two aspects of this problem: 1) for a given erasure code, how to optimally split incoming requests between systematic nodes and repair groups, and 2) choosing an underlying erasure code that maximizes the achievable service capacity region. In particular, we consider MDS and Simplex codes. Our analysis demonstrates that erasure coding makes the system more robust to skews in file popularity than simply replicating a file at multiple servers, and that coding and replication together can make the capacity region larger than either alone.

cs.IT↗

arXiv2017

Effective Straggler Mitigation: Which Clones Should Attack and When?

Redundancy for straggler mitigation, originally in data download and more recently in distributed computing context, has been shown to be effective both in theory and practice. Analysis of systems with redundancy has drawn significant attention and numerous papers have studied pain and gain of redundancy under various service models and assumptions on the straggler characteristics. We here present a cost (pain) vs. latency (gain) analysis of using simple replication or erasure coding for straggler mitigation in executing jobs with many tasks. We quantify the effect of the tail of task execution times and discuss tail heaviness as a decisive parameter for the cost and latency of using redundancy. Specifically, we find that coded redundancy achieves better cost vs. latency and allows for greater achievable latency and cost tradeoff region compared to replication and can yield reduction in both cost and latency under less heavy tailed execution times. We show that delaying redundancy is not effective in reducing cost.

cs.PF↗