Please use this identifier to cite or link to this item: http://hdl.handle.net/10397/85767
Title: A measurement-based approach to support real-time object-based applications timeliness
Authors: Lo, Chun-tat
Degree: M.Phil.
Issue Date: 2006
Abstract: In this paper the novel, original transfer policy framework, namely, the statistical distribution independent transfer policy model (SDITPM) for making sound migration decision in a real-time-manner, is proposed. This framework, which can be incorporated as part of a logical entity (e.g. a logical server) construct, steers the latter effectively in the migration process over a sizeable network such as the Internet. As a result it relieves network congestion and shortens round-trip time through load balancing, which comes naturally from the guided object mobility. The SDITPM framework is a significant contribution to time-critical applications over the Internet because of the shortened service roundtrip time (RTT) in the client/server interactions. Applications for the Internet are intrinsically object-based and exploit the potential speedup supported by the distributed hardware in the underlying network. The objects interact in an end-to-end many-clients-to-one-server relationship called the asymmetric rendezvous via logical channels. If the service RTT for every client/server interaction is shortened, then the overall execution time of the application will be reduced. Therefore, the SDITPM is immensely useful in helping the harnessing of service RTT and making the Internet more usable for real-time computing. The SDITPM mechanism leverages system metrics on the fly, independent of their distribution types (e.g. long-range dependence (LRD) and short-range dependence (SRD)). That is, the SDITPM framework works by direct data measurement statistically, supported by the Convergence Algorithm (CA), which is derived form the Central Limit Theorem. Every system metric being leveraged is called a primary metric, from which the secondary metrics are derived. The present SDITPM framework is basically a FED or "P+I+D" controller (P for proportional control, I for integral control and D for derivative control). The P and D control elements are secondary metrics and the I control sums their control effects for the specific rth control region. In the SDITPM context every primary metric represents a control plane of four regions. Three of them are dedicated separately to PID, PI and DI controls, and the fourth one is inert. From every leveraged primary metric the SDITPM mechanism derives two secondary ones. For example, if the metric is the logical server's queue length Q, then for the rth control region the following are true, a) the P control is the change in Q between two time points (e.g. t1 and t2) and b) the D control is the corresponding rate of change, dQ/dt = (Qt2-Qt1)/(t2-t1) . The control plane/matrix of the primary metric is actually defined by two objective Junctions, {0P,AP}2 for the P control (rows in matrix) and {0D, AD}2 for the D control (columns in the same matrix). The four control regions defined by theses two objective functions are: a) Region 1 (i.e. r1 or rPID) defined by the [P+,D+] pair of positive values for PID control, b) Region 2 (i.e. r2 or rDI) defined by [P-,D+] for "D+I" or DI control, c) Region 3 (i.e. r3 or rPI) defined by [P+,D-] for PI control, and Region 4 (i.e. r4 or rInert) defined by [P-,D-] as an inert or "don't care" state. The regional control (i.e. rth ) only takes the positive value(s) of P and D (i.e. P+ and D+) into account. When the effects of the D and P control are added for the regional control the summation is the integral control effect. In the SDITPM context summation is called the planar integration if only a single metric is being leveraged. If n metrics are leveraged simultaneously, the current control effect for a region rnx is Crnx = nEi=1 (Pix, Dix) where x indicates the control region (e.g. rI or rPID). (Pix, Dix) represents the P and D deviation errors for the objective functions: {0P,AP}2 and {0D, AD}2 respectively. The deviation error [psi] is the difference between the sampled value S and the given safety margin A (e.g. A for P control is AP ) conceptually; [psi] = | S - A | implies [psi] = | Sp - Ap | and [psi]D - | SD - AD | . The emphasis is that the SDITPM framework should leverage only the following conditions: a) positive S and b) S > A for formulating its P, D and I control elements. The migration decision depends on the dominant region, which should currently have the highest Crnx value. Which control region is dominant is totally decided by the system dynamics. Crnx is called the regional transfer probability TPr (i.e. TPr = Crnx), and the threshold for the region is Thr. The dominant TPr becomes the overall transfer probability TPO and Thr of the same region is the overall Th0 threshold. A migration decision at the transfer policy level is affirmed for the condition TPO > ThO. Whether the migration decision will be realized or not depends on the availability of a suitable target node for the logical server to migrate to. Finding the target node is the decision of the location policy, which is not within the scope of the present research. In the SDITPM verification exercise, however, the effect of the location policy is produced by using the A4 visualization package, which is also part of the original contribution by this thesis.
Subjects: Hong Kong Polytechnic University -- Dissertations.
Real-time control.
Real-time data processing.
Electronic data processing -- Distributed processing.
Distributed operating systems (Computers)
Pages: 131 leaves : ill. ; 30 cm.
Appears in Collections:Thesis

Show full item record

Page views

38
Last Week
0
Last month
Citations as of Mar 24, 2024

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.