2009 Workshops at the Grid and Pervasive Computing Conference

Transfer Speed Estimation for Adaptive Scheduling in the Data Grid Mikhail Panshenskov, Alexander Vakhitov Saint Petersburg State University [email protected],[email protected]

Abstract

mance achieved by the scheduler service. Thus by optimizing the analytical service we positively impact scheduling service and in the end overall system performance. In this paper we present general approach based on the assumption of linear dependence of the data transfer time on the size of data. This assumption is true in most cases, and in the very uncertain case of an arbitrary network we cannot make more precise assumptions about the traffic. Here we discuss four different methods, namely Least Squares, Recursive Least Squares (with forgetting factor), Kalman Filter and Randomized Least Squares to estimate the available bandwidth for current data transfers. Those methods do not require additional measurements, but can simply observe the transfers taking place during the Grid computations. The analytic service which uses one of those methods would not disturb the work done by the scheduling service, rather consult the scheduling service with recent estimates of the system parameters. Next, section 2 introduces data grid environment: from abstract to specific level. Main principles of intensive data grid environments are discussed in the beginning, while specific Data Grid implementation used for methods comparison is described in the end of section. In section 3 the problem of channel bandwidth parameters estimation in data grid environment is discussed. Authors set the problem and offer 4 different estimation methods. To compare estimation methods in section 4 simulation scenarios are introduced. In specific Data Grid environment every scenario is performed for every considered method. The results are demonstrated and discussed. The last section 5 concludes the article. Resume of estimation methods and the recommendations of their applications are given.

Four methods to estimate available channel bandwidth in Data Grid are described. To compare methods Data Grid was emulated and the special data transfer scenarios were designed. For every scenario the estimates of the performance and bandwidth parameters are constructed using linear filtering techniques based on the Kalman filter, least-squares, recursive least-squares and linear stochastic approximation with randomized inputs. The results demonstrating the estimation for the available bandwidth parameters are provided and explained.

1

Introduction

With the rise of the network capacity it becomes possible to use Grid computing in the parallel applications where significant amount of data is needed to be transferred. The scheduling algorithms for such distributed Data Grid systems need to be aware of the data transfer delays. It becomes important to estimate the time of the data transfers to allocate the tasks to the Grid nodes effectively. High-performance data transfer is among the main functions of Data Grid [1] where huge amount of data is stored, updated and analyzed. The problem of data transfer tasks scheduling is stated and being solved [2]. To schedule data transfer tasks optimally on data grid environment the parameters (e.g. channel bandwidth) have to be estimated accurately. In this paper we address the problem of available bandwidth estimation from passive measurements of the data transfers through the Grid. Our general system architecture contains 2 main components: analytical and scheduling services. The purpose of analytical service is to provide the information needed for the task allocation which is performed by the scheduling service. This information should include prediction of the task execution time for the specific node and the prediction of the data transfer time between two nodes. The more accurate information provided by the analytical service, the better overall task execution perfor978-0-7695-3677-4/09 $25.00 © 2009 IEEE DOI 10.1109/GPC.2009.21

2

Emulated System of Data Grid

When the amount of transferred data is so high that the incoming or outgoing channel of host becomes overloaded, special traffic scheduling system is needed. The main principles of data intensive systems are: 1. Full load of channels. Transfer channel is a resource of high value. Thus scheduling system should not block 50 58

transfer channels. Full load guarantees that important resources do not stand idle and data channel usage is optimal.

existing approaches to this problem are analyzed, their applicability in the Grid systems is then discussed and the new approach is then proposed. The approaches can be divided into two main classes: capacity bandwidth estimation and available bandwidth estimation [3]. They are both related to our problem: if the capacity of the channel is known, it is possible to set the precise upper bound for the estimates of the data transfer speed. The capacity bandwidth changes not so often in time, so it can be assumed to be constant, stored and accessed later when the transfer using the particular channel is needed again. The knowledge of the available bandwidth is temporary, so if we store the current available bandwidth and if we try to use it later, we can get wrong results because of the changes in the environment. The observations of the available bandwidth which last for some sufficient amount of time can provide information about the time-related patterns of the so-called cross traffic, the third party traffic which accounts for the difference between the capacity and the available bandwidths. The available bandwidth should be estimated in the very same moment as the data transfer is needed to be done, and the method for its estimation can be based on the knowledge of the capacity bandwidth. The capacity bandwidth estimation is based on the socalled packet pair technique, firstly described in [4]. This approach was used later in numerous tools which build the estimates of the capacity bandwidth sending the queue of packets of different size [5, 6, 7, 8, 9, 10]. Adding the assumption about the existence of the crosstraffic in the network, authors move to the available bandwidth estimation. Dispersion between consequent data transfer times is considered to be a predictor for the crosstraffic in [11]. In [12] the queue of the packets with growing delays is sent several times, and the differences between the times of sending and of receiving of the corresponding packets in the queue are supposed to predict the crosstraffic. The authors of [13] and [6] propose the idea to decrease the delays between the packets in the queue. During this process, the moment when the transfer time starts to grow rapidly is determined as a signal that the packets stuck in some buffer in the middle of the path. If the capacity of the bottleneck link in the path is known, then Strauss et al. in [14] provide more accurate estimates for the available bandwidth again based on the packet-pair technique. Dovrolis et al. rely on the estimation of the percentiles of the available bandwidth distribution. They argue that the variation of the transfer times is more helpful to estimate the available bandwidth because according to their main idea the variation is low when the link is saturated and is high when is not, and the four factors (traffic load, number of competing flows, rate of competing flows, measurement timescale) affect the variation. The initial paper [4] provides a general framework for

2. Priority Queue. Scheduling system should put in order set of computing task, target machine couples. This is done to achieve optimal scheduling, thus optimal usage of distributed system. 3. Independence of transfers. Different transfers in the channel should not correlate with each other. This is necessary in order to estimate channel bandwidth accurately. As soon as first criteria is met data channel usage is still optimal. T transf er when host is busy transferring data can be compared to T total of computing the distributed task. If these values are close in logarithmic scale Log(T total ) − Log(T transf ers ) < δ channel is considered to be highly loaded and the computation is data intensive. In respect to principles stated above emulated system of Data Grid environment was built by the authors. For a single computing task network nodes in such environment are divided on a host and clients. Host is responsible for performing a task - spreading task input data, managing execution, collecting task output data. Client is responsible for executing a subtask - receiving input, program execution, sending output. Host is passive meaning that it does not look for clients, on contrary client sends request to host asking for a subtask to execute. Full cycle of host-client interaction looks as follows: 1. (Client→Host) Initialization. Client sends request to host asking for a subtask. 2. (Client←Host) Transfer of Input Data. Host responses with the Input Data for subtask to execute. 3. (Client) Execution. Client executes the subtask. 4. (Client→Host) Transfer Output Data. Client sends the result of execution back to host. 5. (Client←Host) Finalization. Host approves the end of interaction or jumps to step 2. On step 2 and 4 estimation of channel bandwidth is performed.

3

Data Transfer. Available Bandwidth Estimation

The problem is to estimate the transfer time of a data of arbitrary size between two nodes A and B. In this section 59 51

the packet transfer analysis and the time prediction, however the most of the authors use this framework to actively send packets or packet queues. However, those active packet streams need time and substantial channel capacity to be transferred and worked out. Therefore, Seshan et al. [15] propose an architecture of bandwidth estimation based on the passive measurements done by observing the thirdparty traffic. Next, we list the important properties of the networks and data transfers we would like to consider, related to the specificity of the Grid computing and the framework used which was described in the previous section.

sider data tranfer process completely transparent to us making single assumption about its nature. Namely, we assume only that the time of transfer linearly depends on the size of transferring data, which is proved to be true by practical evidence from the GridFTP transfers in [18]. The exact path of a packet is not modeled. We propose to use linear estimation algorithms, namely least squares, recursive least squares, Kalman filtering [16] and simultaneous perturbation [17] to estimate the time of transfer. We make several simulations and show the results for each of the methods discussed.

3.1

1. Grid computations are usually organized over already existing networks which are often used for other purposes as well. Therefore, substantial cross-traffic usually exists. The size and change patterns of the cross traffic are unknown. This traffic is considered as noise in the model which follows.

The problem setting

Let us assume that the data packet of size si is transferred with the time ti . Let us denote cti as a delay in the transfer related to the cross traffic which is unknown. The speed of data transfer through the whole path is supposed to be slightly changing during the observations and is denoted as θi at the moment i. The time needed to open and close the connection or for other purposes which does not depend on si and nevertheless is constantly added to the transfer time we denote as oci . Then:

2. If the third-party network users change the way they transfer information through the channel used by the Grid system, then the available bandwidth changes. Therefore the estimates need not only to converge to true values, but also to adapt to the changing environment. In our linear estimation model (see below) there are well developed methods with theoretical convergence results to deal with such a problem [16, 17].

ti = si ∗ θi + oci + cti .

(1)

The problem can be rewritten for further simplicity as follows: θ˜i = (θi oci )T ; s˜i = (si 1)T . Then all the parameters are gathered in one vector θ˜ which will be further denoted simply as θ. It can be assumed that cti is random, then ti is random as well and the mathematical expectation sign is applicable to them. It can be assumed that si is random as well, which will be important for the fourth algorithm presented here. All the values ti , si , θ, cti ∈ R, and ti , si , θi ≥ 0.

3. Grid usually connects universities in different countries or continents. Because of that, TCP/IP packets usually pass several dozens of intermediary hosts before they reach their destinations. Each link between two hosts can have different bandwidth and cross traffic volumes. The available bandwidth of the whole path is defined as a minimum among the available bandwidths in each of the links. The change in available bandwidth in each of the links can lead to a change in the available bandwidth of the whole path. The number of the intermediary hosts is unknown.

3.2

Estimation algorithms

We use a formulation of the three following algorithms from a conceptual paper of L. Gao [16]. The general form of these algorithms using the notation above is:

4. The packets which are sent in a common session can have different paths in the network. The decisions about the path are made not by the source or destination of the packets, but by intermediary parties such as routers [5]. Therefore it is unfair to rely on a model of a packet path and determine the parameters of each of the links. The path chose by the packet sent can be different from the one we suppose, and the rule of choosing the path depends on the third party routers and is therefore unknown. It cannot even be supposed that this rule has some statistical nature.

θˆi+1 = θˆi + Li (ti − si θˆi ),

(2)

where θˆi is the estimate of the θi , Li is the adaptation gain chosen according to the one of the algorithms presented below. 3.2.1

Because of the last two properties, we propose to use so-called grey-box control in this case. Hereafter we con-

Least Squares [16] Li = μ

60 52

si , 1 + |si |2

(3)

where μ ∈ (0, 1] is called step size or adaptation rate. The algorithm iteratively minimizes the function: Vk (θ) = E(ti − si θi )2 ,

Table 1. Belgium case: mean linear estimation error, ms

(4) Algorithm LS RLS KF Randomized LS

where E is mathematical expectation. 3.2.2

Recursive Least Squares [16] Pi si , α + Pi |si |2 2 2 1 Pi si = (Pi − ), α α + Pi s2i

Scenario 1 319 210 201.5

Scenario 2 1063 649.7 760.2

Scenario 3 431 293 311 317

Li =

Pi+1

Table 2. Saint Petersburg case: mean linear estimation error, ms

(5)

where P0 > 0 and α ∈ (0, 1) is called forgetting factor. The algorithm is derived from minimizing the following criterion: i 1  i−k α (tk − sk θk ). (6) Vi (α) = i

Algorithm LS RLS KF Randomized LS

Scenario 1 942 585 597

Scenario 2 821 650 604

Scenario 3 686 566 595 595

3.2.3

2. Hundred data transfers of files of size gradually increasing from 100KB to 10MB with a step of 100KB,

k=0

Kalman Filtering [16] Pi si , R + Pi s2i Pi2 s2i = Pi − + Q, R + Pi s2i

3. Hundred data transfers of files of size randomly uniformly distributed with values from the set

Li =

Pi+1

{i ∗ 100KB|i = [1..100]}

(7)

There were two experiments conducted. The results of the experiments are shown in the tables 1 and 2. Initial value for all the algorithms was chosen as θˆ0 = (2000, 0)T . The step size parameters (α, μ) were chosen from the interval (0, 1), the one which gives the best results for every particular algorithm. The special Kalman filter parameters were chosen in this manner as well, according to the noise levels. From the tables we can generally conclude that the Recursive Least Squares gives the most reliable results. Kalman Filter method gives almost identical results to the Recursive Least Squares in the available scenarios. The Least Squares method is worse performing in comparison with the other methods, however it still produces reasonable results. To compare methods linear metrics was employed, thus mean linear estimation error was calculated instead of usual mean square estimation error. The cause is that the task scheduler in Grid is likely to summarize estimations for every computing resource. Overall the conducted experiments demonstrated good results — low level error for all four methods utilized. With a time of transfer of 106 bytes of approximately 20 seconds in the both experiments, the average error rate for the scenario 1 was not more than 2% of the transfer time during both experiments and for the scenario 2 it was less than 1%. The quality of the estimates therefore is satisfactory, see

where P0 ≥ 0,R > 0,Q > 0 and θˆ0 are deterministic and can be arbitrarily chosen. R and Q can be regarded as apriori estimates for the variances of cti and θi − θi−1 , respectively. 3.2.4

Randomized Least Squares [17]

Pi+1

Li = Pi Δi , Pi Δi ΔTi Pi = Pi − , 1 + ΔTi Pi Δi Δi = si − Esi ,

(8)

where P0 = γ0−1 I, where I is unit matrix and γ0 > 0. Here we assume that si is chosen from some distribution with finite mean Esi , and all the moments of it are finite.

4

Simulation Scenarios and Results

To represent more accurately real scenario use case, several data transfer scenarios are simulated: 1. Hundred data transfers split on four groups: 0.5MB, 1MB, 1.5MB, 2MB file sizes, 61 53

Fig. 1. From the graph it is evident that the Least Squares method can be good for the comparatively low noise level because it adjusts its estimates quite fast, while the Recursive Least Squares estimates reach the expected value fast and then remain there. The Kalman Filter responses to even small changes in the channel bandwidth. In the Fig. 2 you can see that as the data transfer size grows, linear approximation for the transfer speed becomes more accurate. It is also obvious that the RLS method adapts to big variations in the measurements comparatively fast. The case presented it the Fig. 3 is quite unstable, however the algorithms keep the estimates near the expected true value. Here you can compare the performance of the Randomized LS method and see that the most of its error comes from the more slow initial stage while at the final iterations it provides quite satisfactory estimates.

2800

2600

2400

2200

2000

1800

1600

1400

1200

1000

0

10

20

30

40

50

60

70

80

90

100

Figure 2. Saint Petersburg, scenario 2. The values ti /si (solid line) and the θi estimates via LS(dot), RLS(dot-dash), KF(dash) and Randomized filter (solid with cross).

2800

2600

2000

2400

1950

2200

1900

2000

1850

1800

1800

1600

1750

1400

1700

1200

1650

1000

1600

1550

0

10

20

30

40

50

60

70

80

90

10

20

30

40

50

60

70

80

90

100

Figure 3. Saint Petersburg, scenario 2. The values ti /si (solid line) and the θi estimates via LS(dot), RLS(dot-dash), KF(dash) and Randomized filter (solid with cross), Randomized LS (with stars).

Figure 1. Saint Petersburg, scenario 1. The values ti /si (solid line) and the θi estimates via LS(dot), RLS(dot-dash), KF(dash) and Randomized filter (solid with cross).

From the practical point of view, the most complicated in adjustment is the Kalman Filter, because it needs several parameters to be known (level of the noise, time-varying behavior of the estimated parameter), on which the estimates’ quality significantly depends. From the other side, the Least Squares method also needs adjustments, because the step size suitable in one case in other case can lead to the divergence of the estimates. Recursive Least Squares is quite stable and provides comparable quality of estimates to the Kalman Filter while requiring almost no additional adjustment. The Randomized Least Squares method is also stable. There are modifications of it using the estimates averaging [17] which do not require any a priori information.

5

0

100

cursive Least Squares and Randomized Linear methods. To analyze these methods correctly in emulated Data Grid use case scenarios were introduced and implemented. Two determined and one undetermined scenarios were executed to transfer data between grid domain located in SaintPetersburg,Russia and domain located in Belgium. According to the results in section 4 method employing Recursive Least Squares is the most accurate in general. Kalman Filtering demontrates good results but to use this method optimally one may need a dynamic mechanism to tune method parameters in respect of grid environment change. However, Randomized Linear method does not require dynamic tunning, is fairly accurate and in cases of severe crosstraffic, may outperform other methods. The accuracy of scheduling service in Data Grid highly depends on the accuracy of predictions done by the analytical service. Our results show that if analytical service uses Resursive Least Squares method, the accuracy of the whole

Conclusion

Four different methods of available bandwidth estimation were compared: Kalman Filtering, Least Squares, Re62 54

system will be the best one. In Data Grid small data transfers are needed mostly for data updates or computations. For this kind of transfers as it is shown in section 4 dispersion of the transfer speed is relatively high. Consequently, accuracy of all possible methods of channel bandwidth estimation is relatively low. Replication operation which is needed to Data Grids [1] means transferring big chunks of data. For this kind of transfer operation dispersion of transfer speed is low, thus the estimation methods are more accurate. Consequently comparatively accurate scheduling methods can be used for data replication in Data Grids.

6

[8] K. Lai and M. Baker. Nettimer: A tool for measuring bottleneck link bandwidth // In Proceedings of the USENIX Symposium on Internet Technologies and Systems, 2001, pp. 123134. [9] C. Dovrolis, P. Ramanathan, and D. Moore. PacketDispersion Techniques and a Capacity Estimation Methodology. // IEEE/ACM Transactions on Networking (TON), 12(6). 2004. pp. 963977 [10] R. Kapoor, L.-J. Chen, L. Lao, M. Gerla, and M.Y. Sanadidi. CapProbe: A Simple and Accurate Capacity Estimation Technique // ACM SIGCOMM Computer Communication Review, 34(4). 2004. pp. 6778

Acknowledgements

[11] R. L. Carter, M. E. Crovella Dynamic server selection using bandwidth probing in wide-area networks // Technical Report. 1996.

This work was supported by grant from Intel Corporation as a project in SPRInt laboratory, Saint Petersburg State University. We thank Prof. Oleg. N. Granichin for assistance with algorithms formulation.

[12] N. Hu and P. Steenkiste. Evaluation and characterization of available bandwidth and probing techniques // IEEE JSAC Special Issue in Internet and WWW Measurement, Mapping, and Modeling, 21. 2003. pp. 879894

References [1] S.Venugopal, R.Buyya, K.Ramamohanarao. A taxonomy of Data Grids for distributed data sharing, management, and processing // ACM Computing Surveys. vol. 38, issue 1. 2006. pp. 1-53.

[13] M. Jain and C. Dovrolis. End-to-End Available Bandwidth : Measurement Methodology, Dynamics, and Relation with TCP Throughput // IEEE/ACM Transactions on Networking, 11(4) 2003. pp. 537549.

[2] K.Ranganathan , I.Foster. Decoupling Computation and Data Scheduling in Distributed Data-Intensive Applications // In proceedings of the 11th IEEE International Symposium on High Performance Distributed Computing, July 24-26, 2002, p.352.

[14] J. Strauss, D. Katabi, and F. Kaashoek. A Measurement Study of Available Bandwidth Estimation Tools. // In ACM/USENIX Internet Measurement Conference (IMC). 2003. pp. 39-44.

[3] M. Neginhal. Efficient Estimation of Available Bandwidth Along Network Paths // Masters Thesis, North Carolina State University, 2006.

[15] S. Seshan, M. Stemm, and R. H. Katz. SPAND: Shared passive network performance discovery // In USENIX Symposium on Internet Technologies and Systems, 1997.

[4] S. Keshav. A Control-Theoretic Approach to Flow Control // ACM SIGCOMM Computer Communication Review. vol. 5, issue 1. 1995. pp. 188-201.

[16] Guo L., “Stability of recursive stochastic tracking algorithms,” SIAM J. Control and Optimization, vol. 32, No 5. 1994. pp. 1195–1225.

[5] V. Paxson. End-to-End Internet Packet Dynamics. // IEEE/ACM Transactions on Networking (TON), 7(3). 1999. pp. 277292

[17] Granichin O. N., “Linear Regression and Filtering Under Nonstandard Assumptions (Arbitrary Noise),” IEEE Trans. Automat. Contr., vol. 44, No 3. 1999. pp. 442–453.

[6] V. Ribeiro, R. Riedi, R. Baraniuk, J. Navratil, and L. Cottrell. pathChirp: Efficient Available Bandwidth Estimation for Network Paths // In Proceedings of The Conference on Passive and Active Measurements (PAM). April 2003.

[18] S. Vazhkudai, J. M. Schopf. Predicting Sporadic Grid Data Transfers // In Proceedings of HPDC-2002. 2002.

[7] R.L. Carter and M.E. Crovella. Measuring Bottleneck Link Speed in Packet-Switched Networks // Performance Evaluation, 27(28). 1996. pp. 297318. 63 55

Transfer Speed Estimation for Adaptive Scheduling in the Data Grid

Saint Petersburg State University [email protected],[email protected]. Abstract. Four methods to estimate available channel bandwidth in Data Grid are ...

303KB Sizes 1 Downloads 253 Views

Recommend Documents

Adaptive Data Block Scheduling for Parallel TCP Streams
TCP [10], Scalable TCP [19], BIC-TCP [30], and Conges- tion Manager [7]. 4. ... turn causes the scheduling control system to reduce the number of packets ...... for high performance, wide-area distributed file downloads,”. Parallel Processing ...

A Scheduling Method for Divisible Workload Problem in Grid ...
ing algorithms. Section 3 briefly describes our hetero- geneous computation platform. Section 4 introduces our dynamic scheduling methodology. Section 5 con-.

accelerometer - enhanced speed estimation for ... - Infoscience - EPFL
have to be connected to the mobile slider part. It contains the ... It deals with design and implementation of controlled mechanical systems. Its importance ...... More precise and cheaper sensors are to be expected in the future. 3.2 Quality of ...

accelerometer - enhanced speed estimation for ... - Infoscience - EPFL
A further increase in position resolution limits the maximum axis speed with today's position encoders. This is not desired and other solutions have to be found.

Adaptive Scheduling Parameters Manager for ... - GitHub
Jun 27, 2014 - Solution: a set of tools that manage SCHED DEADLINE parameters adaptively ..... Adaptive Quality of Service Architecture. Wiley. InterScience ...

Wireless Power Transfer for Distributed Estimation in ...
wireless sensors are equipped with radio-frequency based en- ergy harvesting .... physical-layer security problems for multiuser MISO [24], and multiple-antenna ..... energy beams are required for the optimal solution of problem. (SDR1). ...... Journ

Scheduling Mixed Workloads in Multi-grids: The Grid ...
pools (which we call grids) that vary significantly in their ... tion level for a task is dictated by the task's complexity. .... in any way. ...... In 16th Conference on Un-.

Cheap High Speed Usb 2.0 Data Transfer Cable Blue Color ...
Cheap High Speed Usb 2.0 Data Transfer Cable Blue C ... Play Usb Cable Free Shipping & Wholesale Price.pdf. Cheap High Speed Usb 2.0 Data Transfer ...

Case Study of QoS Based Task Scheduling for Campus Grid
Also Grid computing is a system, which provides distributed services that integrates wide variety of resources with ... communication based jobs are like transfer a file from one system to another system and which require high ... Section 4 is relate

Case Study of QoS Based Task Scheduling for Campus Grid
Such Grids give support to the computational infrastructure. (access to computational and data ... Examples of Enterprise Grids are Sun Grid Engine, IBM. Grid, Oracle Grid' and HP Grid ... can be used. Here m represents the communicational types of t

Adaptive Scheduling for Multicasting Hard Deadline ...
tructure support, scalable compression techniques are used to allow the .... that the transmitter is an oracle which knows the channel ..... global optimal solution.

Adaptive resource allocation and frame scheduling for wireless multi ...
allocation and frame scheduling concept for wireless video streaming. ... I. INTRODUCTION. Wireless multimedia communication is challenging due to the time ...

Noise Plus Interference Power Estimation in Adaptive ...
possible data rates, the transmission bandwidth of OFDM systems is also large ... where Sn,k is the transmitted data symbol at the kth subcarrier of the nth OFDM ...

decentralized set-membership adaptive estimation ... - Semantic Scholar
Jan 21, 2009 - new parameter estimate. Taking advantage of the sparse updates of ..... cursive least-squares using wireless ad hoc sensor networks,”. Proc.

ESTIMATION OF CAUSAL STRUCTURES IN LONGITUDINAL DATA ...
in research to study causal structures in data [1]. Struc- tural equation models [2] and Bayesian networks [3, 4] have been used to analyze causal structures and ...

Estimation of the phase derivative using an adaptive window ...
rivative contains the information regarding the measur- and. During recent years, there has been an increased in- terest in developing methods that can directly estimate the phase derivative from a fringe pattern because the phase derivative conveys

Data Enrichment for Incremental Reach Estimation
12.32. 0.29. Tab. 2.1: Summary statistics for 6 campaigns: web reach in SSP, TV reach in ... marketing questions of interest have to do with aggregates. Using the ...

Performance Evaluation of Grid Scheduling Strategies: A Case ... - IJRIT
IJRIT International Journal of Research in Information Technology, Volume 2, Issue 4, ... problems that are characterized by having high degree of parallelism.

Performance Evaluation of Grid Scheduling Strategies: A Case ... - IJRIT
tasks are needed to identify galaxy clusters within the Sloan Digital Sky Survey [3]. Because of large amounts of computation and data involved, these workflows ...

A Grid-Based Location Estimation Scheme using Hop ...
Report DCS-TR-435, Rutgers University, April 2001. [13] J. G. Lim, K. L. Chee, H. B. Leow, Y. K. Chong, P. K. Sivaprasad and. SV Rao, “Implementing a ...

Data Transfer Project Services
As Download Your Data grows, and as more companies create portability offerings of their own,. Google continues .... The legacy Provider has limited options for.