Automatic Data Allocation and Buffer Management for Multi-GPU Machines

Multi-GPU machines are being increasingly used in high-performance computing. Each GPU in such a machine has its own memory and does not share the address space either with the host CPU or other GPUs. Hence, applications utilizing multiple GPUs have to manually allocate and manage data on each GPU. Existing works that propose to automate data allocations for GPUs have limitations and inefficiencies in terms of allocation sizes, exploiting reuse, transfer costs, and scalability. We propose a scalable and fully automatic data allocation and buffer management scheme for affine loop nests on multi-GPU machines. We call it the Bounding-Box-based Memory Manager (BBMM). BBMM can perform at runtime, during standard set operations like union, intersection, and difference, finding subset and superset relations on hyperrectangular regions of array data (bounding boxes). It uses these operations along with some compiler assistance to identify, allocate, and manage data required by applications in terms of disjoint bounding boxes. This allows it to (1) allocate exactly or nearly as much data as is required by computations running on each GPU, (2) efficiently track buffer allocations and hence maximize data reuse across tiles and minimize data transfer overhead, and (3) and as a result, maximize utilization of the combined memory on multi-GPU machines. BBMM can work with any choice of parallelizing transformations, computation placement, and scheduling schemes, whether static or dynamic. Experiments run on a four-GPU machine with various scientific programs showed that BBMM reduces data allocations on each GPU by up to 75% compared to current allocation schemes, yields performance of at least 88% of manually written code, and allows excellent weak scaling.

On Finding the Natural Number of Topics with Latent Dirichlet Allocation: Some Observations

It is important to identify the ``correct'' number of topics in mechanisms like Latent Dirichlet Allocation(LDA) as they determine the quality of features that are presented as features for classifiers like SVM. In this work we propose a measure to identify the correct number of topics and offer empirical evidence in its favor in terms of classification accuracy and the number of topics that are naturally present in the corpus. We show the merit of the measure by applying it on real-world as well as synthetic data sets(both text and images). In proposing this measure, we view LDA as a matrix factorization mechanism, wherein a given corpus C is split into two matrix factors M-1 and M-2 as given by C-d*w = M1(d*t) x Q(t*w).Where d is the number of documents present in the corpus anti w is the size of the vocabulary. The quality of the split depends on ``t'', the right number of topics chosen. The measure is computed in terms of symmetric KL-Divergence of salient distributions that are derived from these matrix factors. We observe that the divergence values are higher for non-optimal number of topics - this is shown by a `dip' at the right value for `t'.

Soft classification based Sub-pixel allocation model

Land cover (LC) refers to what is actually present on the ground and provide insights into the underlying solution for improving the conditions of many issues, from water pollution to sustainable economic development. One of the greatest challenges of modeling LC changes using remotely sensed (RS) data is of scale-resolution mismatch: that the spatial resolution of detail is less than what is required, and that this sub-pixel level heterogeneity is important but not readily knowable. However, many pixels consist of a mixture of multiple classes. The solution to mixed pixel problem typically centers on soft classification techniques that are used to estimate the proportion of a certain class within each pixel. However, the spatial distribution of these class components within the pixel remains unknown. This study investigates Orthogonal Subspace Projection - an unmixing technique and uses pixel-swapping algorithm for predicting the spatial distribution of LC at sub-pixel resolution. Both the algorithms are applied on many simulated and actual satellite images for validation. The accuracy on the simulated images is ~100%, while IRS LISS-III and MODIS data show accuracy of 76.6% and 73.02% respectively. This demonstrates the relevance of these techniques for applications such as urban-nonurban, forest-nonforest classification studies etc.

Fuzzy inference based data preprocessing for VBR video traffic prediction

Prediction of variable bit rate compressed video traffic is critical to dynamic allocation of resources in a network. In this paper, we propose a technique for preprocessing the dataset used for training a video traffic predictor. The technique involves identifying the noisy instances in the data using a fuzzy inference system. We focus on three prediction techniques, namely, linear regression, neural network and support vector regression and analyze their performance on H.264 video traces. Our experimental results reveal that data preprocessing greatly improves the performance of linear regression and neural network, but is not effective on support vector regression.

Mining sentiments from songs using latent dirichlet allocation

Song-selection and mood are interdependent. If we capture a song’s sentiment, we can determine the mood of the listener, which can serve as a basis for recommendation systems. Songs are generally classified according to genres, which don’t entirely reflect sentiments. Thus, we require an unsupervised scheme to mine them. Sentiments are classified into either two (positive/negative) or multiple (happy/angry/sad/...) classes, depending on the application. We are interested in analyzing the feelings invoked by a song, involving multi-class sentiments. To mine the hidden sentimental structure behind a song, in terms of “topics”, we consider its lyrics and use Latent Dirichlet Allocation (LDA). Each song is a mixture of moods. Topics mined by LDA can represent moods. Thus we get a scheme of collecting similar-mood songs. For validation, we use a dataset of songs containing 6 moods annotated by users of a particular website.

Obtaining single document summaries using latent dirichlet allocation

In this paper, we present a novel approach that makes use of topic models based on Latent Dirichlet allocation(LDA) for generating single document summaries. Our approach is distinguished from other LDA based approaches in that we identify the summary topics which best describe a given document and only extract sentences from those paragraphs within the document which are highly correlated given the summary topics. This ensures that our summaries always highlight the crux of the document without paying any attention to the grammar and the structure of the documents. Finally, we evaluate our summaries on the DUC 2002 Single document summarization data corpus using ROUGE measures. Our summaries had higher ROUGE values and better semantic similarity with the documents than the DUC summaries.

Training for Antenna Selection in Time-Varying Channels: Optimal Selection, Energy Allocation, and Energy Efficiency Evaluation

Single receive antenna selection (AS) is a popular method for obtaining diversity benefits without the additional costs of multiple radio receiver chains. Since only one antenna receives at any time, the transmitter sends a pilot multiple times to enable the receiver to estimate the channel gains of its N antennas to the transmitter and select an antenna. In time-varying channels, the channel estimates of different antennas are outdated to different extents. We analyze the symbol error probability (SEP) in time-varying channels of the N-pilot and (N+1)-pilot AS training schemes. In the former, the transmitter sends one pilot for each receive antenna. In the latter, the transmitter sends one additional pilot that helps sample the channel fading process of the selected antenna twice. We present several new results about the SEP, optimal energy allocation across pilots and data, and optimal selection rule in time-varying channels for the two schemes. We show that due to the unique nature of AS, the (N+1)-pilot scheme, despite its longer training duration, is much more energy-efficient than the conventional N-pilot scheme. An extension to a practical scenario where all data symbols of a packet are received by the same antenna is also investigated.

A Markov Decision Theoretic Approach to Pilot Allocation and Receive Antenna Selection

This paper considers antenna selection (AS) at a receiver equipped with multiple antenna elements but only a single radio frequency chain for packet reception. As information about the channel state is acquired using training symbols (pilots), the receiver makes its AS decisions based on noisy channel estimates. Additional information that can be exploited for AS includes the time-correlation of the wireless channel and the results of the link-layer error checks upon receiving the data packets. In this scenario, the task of the receiver is to sequentially select (a) the pilot symbol allocation, i.e., how to distribute the available pilot symbols among the antenna elements, for channel estimation on each of the receive antennas; and (b) the antenna to be used for data packet reception. The goal is to maximize the expected throughput, based on the past history of allocation and selection decisions, and the corresponding noisy channel estimates and error check results. Since the channel state is only partially observed through the noisy pilots and the error checks, the joint problem of pilot allocation and AS is modeled as a partially observed Markov decision process (POMDP). The solution to the POMDP yields the policy that maximizes the long-term expected throughput. Using the Finite State Markov Chain (FSMC) model for the wireless channel, the performance of the POMDP solution is compared with that of other existing schemes, and it is illustrated through numerical evaluation that the POMDP solution significantly outperforms them.

Subchannel allocation and power control in femtocells to provide quality of service

Femtocells are a new concept which improves the coverage and capacity of a cellular system. We consider the problem of channel allocation and power control to different users within a Femtocell. Knowing the channels available, the channel states and the rate requirements of different users the Femtocell base station (FBS), allocates the channels to different users to satisfy their requirements. Also, the Femtocell should use minimal power so as to cause least interference to its neighboring Femtocells and outside users. We develop efficient, low complexity algorithms which can be used online by the Femtocell. The users may want to transmit data or voice. We compare our algorithms with the optimal solutions.

Optimal Power Allocation Policies for Low Powered Wireless Communication Systems

We consider optimal average power allocation policies in a wireless channel in the presence of individual delay constraints on the transmitted packets. Power is consumed in transmission of data only. We consider the case when the power used in transmission is a linear function of the data transmitted. The transmission channel may experience multipath fading. We have developed a computationally efficient online algorithm, when there is same hard delay constraint for all packets. Later on, we generalize it to the case when there are multiple real time streams with different hard deadline constraints. Our algorithm uses linear programming and has very low complexity.

QEP Based Solution for Power Allocation in Amplify and Forward Networks

The problem of cooperative beamforming for maximizing the achievable data rate of an energy constrained two-hop amplify-and-forward (AF) network is considered. Assuming perfect channel state information (CSI) of all the nodes, we evaluate the optimal scaling factor for the relay nodes. Along with individual power constraint on each of the relay nodes, we consider a weighted sum power constraint. The proposed iterative algorithm initially solves a set of relaxed problems with weighted sum power constraint and then updates the solution to accommodate individual constraints. These relaxed problems in turn are solved using a sequence of Quadratic Eigenvalue Problems (QEP). The key contribution of this letter is the generalization of cooperative beamforming to incorporate both the individual and weighted sum constraint. Furthermore, we have proposed a novel algorithm based on Quadratic Eigenvalue Problem (QEP) and discussed its convergence.

Efficient Low Complexity Power Allocation Policies for Wireless Communication Systems Guaranteeing QoS

We consider near-optimal policies for a single user transmitting on a wireless channel which minimize average queue length under average power constraint. The power is consumed in transmission of data only. We consider the case when the power used in transmission is a linear function of the data transmitted. The transmission channel may experience multipath fading. Later, we also extend these results to the multiuser case. We show that our policies can be used in a system with energy harvesting sources at the transmitter. Next we consider data users which require minimum rate guarantees. Finally we consider the system which has both data and real time users. Our policies have low computational complexity, closed form expression for mean delays and require only the mean arrival rate with no queue length information.