Several recently devised machine learning (ML) algorithms have shown improved accuracy for various predictive problems. Model searches, which explore to find an optimal ML algorithm and hyperparameter values for the target problem, play a critical role in such improvements. During a model search, data scientists typically use multiple ML implementations to construct several predictive models; however, it takes significant time and effort to employ multiple ML implementations due to the need to learn how to use them, prepare input data in several different formats, and compare their outputs. Our proposed framework addresses these issues by providing simple and unified coding method. It has been designed with the following two attractive features: i) new machine learning implementations can be added easily via common interfaces between the framework and ML implementations and ii) it can be scaled to handle large model configuration search spaces via profile-based scheduling. The results of our evaluation indicate that, with our framework, implementers need only write 55-144 lines of code to add a new ML implementation. They also show that ours was the fastest framework for the HIGGS dataset, and the second-fastest for the SECOM dataset.

The importance of interpretability of machine learning models has been increasing due to emerging enterprise predictive analytics. Piecewise linear models have been actively studied to achieve both accuracy and interpretability. They often produce competitive accuracy against state-of-the-art non-linear methods. In addition, their representations (i.e., rule-based segmentation plus sparse linear formula) are often preferred by domain experts. A disadvantage of such models, however, is high computational cost for simultaneous determinations of the number of “pieces” and cardinality of each linear predictor, which has restricted their applicability to middle-scale data sets. This paper discusses an empirical study on the derivation of a distributed factorized asymptotic Bayesian (FAB) inference of learning piece-wise sparse linear models on distributed memory architectures from the original FAB inference algorithm. The distributed FAB inference solves the simultaneous model selection issue without communicating O(N) data where N is the number of training samples and achieves linear scale-out against the number of CPU cores. Experimental results demonstrate that the distributed FAB inference achieves high prediction accuracy and performance scalability with both synthetic and public benchmark data.

This paper proposes a distributed factorized asymptotic Bayesian (FAB) inference of learning piece-wise sparse linear models on distributed memory architectures. The distributed FAB inference solves the simultaneous model selection issue without communicating O(N) data where N is the number of training samples and achieves linear scale-out against the number of CPU cores. It unlocks the limitation of their applicability to middle-scale data sets due to high computational cost for simultaneous determinations of the number of “pieces” and cardinality of each linear predictor. Experimental results demonstrate that the distributed FAB inference achieves high prediction accuracy and performance scalability with benchmark data.

Cloud-federations have emerged as popular platforms for Internet-scale services. Cloud-federations are running over multiple datacenters, because a cloud-federation is an aggregate of cloud services each of which runs in a single datacenter. In such inter-datacenter environments, distributed key-value stores (DKVSs) are attractive databases in terms of scalability. However, inter-datacenter communications degrade the performance of these DKVSs because of their large latency and narrow bandwidth. In this paper, we demonstrate how to reduce and hide the weak points of inter-datacenter communications for DKVSs. To solve the problems we introduce two techniques called multi-layered DHT (ML-DHT) and local-first data rebuilding (LDR). ML-DHT provides a global and consistent index of key-value pairs with the efficient expandability of the storage capacity. It employs a routing protocol which reduces routing hops that pass through interdatacenter connections. LDR reduces data transfer on interdatacenter connections by using erasure coding techniques. It enables KVS administrators to flexibly make trade-offs between expandability of storage capacity and the performance of data transfer. Experimental results demonstrate that our techniques improve the latency up to 74 % compared with a Chord-based system and enable us to balance the amount of storage usage and remote data transfer.

This paper presents Pangaea, an inter-data center key-value store that keeps reasonable expandability of storage capacity, data lookup latency and data transfer speed. Pangaea uses two techniques called multi-layered DHT (ML-DHT) and local-first data rebuilding (LDR). ML-DHT provides a global and consistent index of key-value pairs with efficient routings in inter-data center environments. LDR reduces inter-data center data transfer by using erasure coding techniques.

Data-intensive computing systems like MapReduce and Dryad have emerged as a framework for leveraging computing resources of a cluster. I/O bottlenecks need to be eased to improve performance in data-intensive computing systems. State-of-the-art frameworks for data-intensive computing have tackled the issue with a data locality based task scheduling policy. However, locality-aware scheduling does not always work good to mitigate I/O bottlenecks when different I/O characteristic jobs run concurrently. This paper presents LoadAtomizer, a locality and I/O load aware task scheduler for MapReduce. LoadAtomizer mitigates the I/O bottlenecks of a cluster with locality and I/O load aware map task assignment and storage selection. LoadAtomizer quickly assigns a slave a map task whose input data is stored in a lightly loaded storage and commands the slave to read the input data from the storage. LoadAtomizer maintains the load information of storages and the network with a topology-aware load tree. A topology-aware load tree enables LoadAtomizer to select quickly a lightly loaded storage that a slave can access through a lightly loaded network path. Experimental results demonstrated that our prototype of LoadAtomizer shortened completion time of multiple jobs by up to 18.6 %.

As demand for high fidelity multimedia content has soared, content distribution has emerged as a critical application. Large multimedia files require effective content distribution services such as content distribution networks (CDNs). A recent trend in CDN development is the use of peer-to-peer (P2P) techniques to enhance scalability, fault resilience, and cost-effectiveness. Unfortunately, P2P-based content distribution poses a crucial problem in that update propagation is quite difficult to accomplish. This is because peers cannot obtain a global view of replica locations on the network. In this paper, we propose speculative update, which quickly propagates an update to all replicas in a pure P2P fashion. Each server attempts to determine the directions in which there will be replicas with high probability based on server's local state used for replica repositioning. Then, it relays update messages speculatively in those directions. Simulation results demonstrate that our mechanism propagates an update to all replicas faster than the current pure P2P-based approaches.

Network coordinates (NCs) construct a logical space which enables efficient and accurate estimation of network latency. Although many researchers have proposed NC-based strategies to reduce the lookup latency of distributed hash tables (DHTs), these strategies are limited in the improvement of the lookup latency; the nearest node to which a query should be forwarded is not always included in the consideration scope of a node. This is because conventional DHTs assign node IDs independent of the underlying physical network. In this paper, we propose an NC-based method of constructing a topology-aware DHT by Proximity Identifier Selection strategy (PIS/NC). PIS/NC assigns an ID to each node based on NC of the node. This paper presents Canary, a PIS/NC-based CAN whose d-dimensional logical space corresponds to that of Vivaldi. Our simulation results suggest that PIS/NC has the possibility of dramatically improving the lookup latency of DHTs. Whereas DHash++ is only able to reduce the median lookup latency by 15% of the original Chord, Canary reduces it by 70% of the original CAN.

Many service providers distribute various kinds of content over the Internet. They often use replica servers to provide stable service. To position them appropriately, service providers must predict the demands for their services and provide computing capacity sufficient for servicing the demands. Unfortunately, predicting demands is difficult because demand for a service usually fluctuates. Our research group is developing ExaPeer, an infrastructure that apportions computing capacity to services running on hundreds or thousands of trusted machines all over the Internet. In this paper, we describe ExaPeer’s approach to dynamically selecting candidate spots for replica servers. The runtime system in ExaPeer detects fluctuations in demand and then dynamically selects candidate spots on which replica servers should be placed to best meet the demand. Experimental results demonstrate that the candidate spots selected by ExaPeer work better than manually selected ones even if the scale of demand changes rapidly.