Luận án Nghiên cứu phát triển một số kỹ thuật hỗ trợ phát hiện đạo văn và ứng dụng cho văn bản Tiếng Việt

e time. Our approach to addressing the networking requirements for live WAN migration builds on the observations that not all networking changes in this approach are time critical and further that instantaneous changes are best achieved in a localized manner. Specifically, in our solution, described in detail in Section 3, we allow the migration software to initiate the necessary networking changes as soon as the need for migration has been identified. We make use of tunneling technologies during this initial phase to preemptively establish connectivity between the data centers involved. Once server migration is complete, the migration software initiates a local change to direct traffic towards the new data center via the tunnel. Slower time scale network changes then phase out this local network connectivity change for a more optimal network wide path to the new data center. 2.3 Storage Replication Requirements Data availability is typically addressed by replicating business data on a local/primary storage system, to some remote location from where it can be accessed. From a business/usability point of view, such remote replication is driven by two metrics [9]. First 263 is the recovery-point-objective which is the consistent data point to which data can be restored after a disaster. Second is the recoverytime-objective which is the time it takes to recover to that consistent data point after a disaster [13]. Remote replication can be broadly classified into the following two categories: ¡ Synchronous replication: every data block written to a local P.6 storage system is replicated to the remote location before the local write operation returns. ¡ Asynchronous replication: in this case the local and remote storage systems are allowed to diverge. The amount of divergence between the local and remote copies is typically bounded by either a certain amount of data, or by a certain amount of time. Synchronous replication is normally recommended for applications, such as financial databases, where consistency between local and remote storage systems is a high priority. However, these desirable properties come at a price. First, because every data block needs to be replicated remotely, synchronous replication systems can not benefit from any local write coalescing of data if the same data blocks are written repeatedly [16]. Second, because data have to be copied to the remote location before the write operation returns, synchronous replication has a direct performance impact on the application, since both lower throughput and increased latency of the path between the primary and the remote systems are reflected in the time it takes for the local disk write to complete. An alternative is to use asynchronous replication. However, because the local and remote systems are allowed to diverge, asynchronous replication always involves some data loss in the event of a failure of the primary system. But, because write operations can be batched and pipelined, asynchronous replication systems can move data across the network in a much more efficient manner than synchronous replication systems. For WAN live server migration we seek a more flexible replication system where the mode can be dictated by the migration semantics. Specifically, to support live server migration we propose a remote replication system where the initial transfer of data between the data centers is performed via asynchronous replication to benefit from the efficiency of that mode of operation. When the bulk of the data have been transfered in this manner, replication switches to synchronous replication in anticipation of the completion of the server migration step. The final server migration step triggers a simultaneous switch-over to the storage system at the new data center. In this manner, when the virtual server starts executing in the new data center, storage requirements can be locally met. 3. WAN MIGRATION SCENARIOS In this section we illustrate how our cooperative, context aware approach can combine the technical building blocks described in the previous section to realize live server migration across a wide area network. We demonstrate how the coordination of server virtualization and migration technologies, the storage replication P.7 subsystem and the network can achieve live migration of the entire data center across the WAN. We utilize different scenarios to demonstrate our approach. In Section 3.1 we outline how our approach can be used to achieve the safe live migration of a data center when planned maintenance events are handled. In Section 3.2 we show the use of live server migration to mitigate the effects of unplanned outages or failures. 3.1 Maintenance Outages We deal with maintenance outages in two parts. First, we consider the case where the service has no (or very limited) storage requirements. This might for example be the case with a network element such as a voice-over-IP (VoIP) gateway. Second, we deal with the more general case where the service also requires the migration of data storage to the new data center. Without Requiring Storage to be Migrated: Without storage to be replicated, the primary components that we need to coordinate are the server migration and network mobility. Figure 1 shows the environment where the application running in a virtual server VS has to be moved from a physical server in data center A to a physical server in data center B. Prior to the maintenance event, the coordinating migration management system (MMS) would signal to both the server management system as well as the network that a migration is imminent. The server management system would initiate the migration of the virtual server from physical server a (¢¤£¦¥ ) to physical server b (¢¤£¦§ ). After an initial bulk state transfer as preparation for migration, the server management system will mirror any state changes between the two virtual servers. Similarly, for the network part, based on the signal received from the MMS, the service provider edge (¢©¨ ) router will initiate a number of steps to prepare for the migration. Specifically, as shown in Figure 1(b), the migration system will cause the network to create a tunnel between ¢©¨and ¢©¨which will be used subsequently to transfer data destined to VS to data center B. When the MMS determines a convenient point to quiesce the VS, another signal is sent to both the server management system and the network. For the server management system, this signal will indicate the final migration of the VS from data center A to data center B, i.e., after this the VS will become active in data center B. For the network, this second signal enables the network data path to switchover locally at ¢©¨©¥ to the remote data center. Specifically, from this point in time, any traffic destined for the virtual server address that arrives at ¢©¨©¥ will be switched onto the tunnel to ¢©¨©§ for delivery to data center B. P.8 Note that at this point, from a server perspective the migration is complete as the VS is now active in data center B. However, traffic is sub-optimally flowing first to ¢©¨©¥ and then across the tunnel to ¢©¨¤§ . To rectify this situation another networking step is involved. Specifically, ¢©¨©§ starts to advertise a more preferred route to reach VS, than the route currently being advertised by ¢©¨¤¥ . In this manner, as ingress PEs to the network (¢©¨¤to ¢©¨¤ in Figure 1) receive the more preferred route, traffic will start to flow to ¢©¨©§ directly and the tunnel between ¢©¨©¥ and ¢©¨©§ can be torn down leading to the final state shown in Figure 1(c). Requiring Storage Migration: When storage has to also be replicated, it is critical that we achieve the right balance between performance (impact on the application) and the recovery point or data loss when the switchover occurs to the remote data center. To achieve this, we allow the storage to be replicated asynchronously, prior to any initiation of the maintenance event, or, assuming the amount of data to be transfered is relatively small, asynchronous replication can be started in anticipation of a migration that is expected to happen shortly. Asynchronous replication during this initial phase allows for the application to see no performance impact. However, when the maintenance event is imminent, the MMS would signal to the replication system to switch from asynchronous replication to synchronous replication to ensure that there is no loss of data during migration. When data is being replicated synchronously, there will be a performance impact on the application. 264 Figure 1: Live server migration across a WAN This requires us to keep the exposure to the amount of time we replicate on a synchronous basis to a minimum. When the MMS signals to the storage system the requirement to switch to synchronous replication, the storage system completes all the pending asynchronous operations and then proceeds to perform all the subsequent writes by synchronously replicating it to the remote data center. Thus, between the server migration and synchronous replication, both the application state and all the storage operations are mirrored at the two environments in the two data centers. When all the pending write operations are copied over, then as in the previous case, we quiesce the application and the network is signaled to switch traffic over to the remote data center. From this point on, both storage and server migration operations are complete and activated in data center B. As above, the network state still needs to be updated to ensure optimal data flow directly to data center B. Note that while we have described the live server migration P.9 process as involving the service provider for the networking part, it is possible for a data center provider to perform a similar set of functions without involving the service provider. Specifically, by creating a tunnel between the customer edge (CE) routers in the data center, and performing local switching on the appropriate CE, rather than on the PE, the data center provider can realize the same functionality. 3.2 Unplanned Outages We propose to also use cooperative, context aware migration to deal with unplanned data center outages. There are multiple considerations that go into managing data center operations to plan and overcome failures through migration. Some of these are: (1) amount of overhead under normal operation to overcome anticipated failures; (2) amount of data loss affordable (recovery point objective - RPO); (3) amount of state that has to be migrated; and (4) time available from anticipated failure to occurrence of event. At the one extreme, one might incur the overhead of completely mirroring the application at the remote site. This has the consequence of both incurring processing and network overhead under normal operation as well as impacting application performance (latency and throughput) throughout. The other extreme is to only ensure data recovery and to start a new copy of the application at the remote site after an outage. In this case, application memory state such as ongoing sessions are lost, but data stored on disk is replicated and available in a consistent state. Neither this hot standby nor the cold standby approach described are desirable due to the overhead or the loss of application memory state. An intermediate approach is to recover control and essential state of the application, in addition to data stored on disk, to further minimize disruptions to users. A spectrum of approaches are possible. In a VoIP server, for instance, session-based information can be mirrored without mirroring the data flowing through each session. More generally, this points to the need to checkpoint some application state in addition to mirroring data on disk. Checkpointing application state involves storing application state either periodically or in an application-aware manner like databases do and then copying it to the remote site. Of course, this has the consequence that the application can be restarted remotely at the checkpoint boundary only. Similarly, for storage one may use asynchronous replication with a periodic snapshot ensuring all writes are up-to-date at the remote site at the time of checkpointing. Some data loss may occur upon an unanticipated, catastrophic failure, but the recovery point may be fairly small, depending on the frequency of checkpointing application and storage state. Coordination between P.10 265 the checkpointing of the application state and the snapshot of storage is key to successful migration while meeting the desired RPOs. Incremental checkpointing of application and storage is key to efficiency, and we see existing techniques to achieve this [4, 3, 11]. For instance, rather than full application mirroring, a virtualized replica can be maintained as a warm standby-in dormant or hibernating state-enabling a quick switch-over to the previously checkpointed state. To make the switch-over seamless, in addition to replicating data and recovering state, network support is needed. Specifically, on detecting the unavailability of the primary site, the secondary site is made active, and the same mechanism described in Section 3.1 is used to switch traffic over to reach the secondary site via the pre-established tunnel. Note that for simplicity of exposition we assume here that the PE that performs the local switch over is not affected by the failure. The approach can however, easily be extended to make use of a switchover at a router deeper in the network. The amount of state and storage that has to be migrated may vary widely from application to application. There may be many situations where, in principle, the server can be stateless. For example, a SIP proxy server may not have any persistent state and the communication between the clients and the proxy server may be using UDP. In such a case, the primary activity to be performed is in the network to move the communication over to the new data center site. Little or no overhead is incurred under normal operation to enable the migration to a new data center. Failure recovery involves no data loss and we can deal with near instantaneous, catastrophic failures. As more and more state is involved with the server, more overhead is incurred to checkpoint application state and potentially to take storage snapshots, either periodically or upon application prompting. It also means that the RPO is a function of the interval between checkpoints, when we have to deal with instantaneous failures. The more advanced information we have of an impending failure, the more effective we can be in having the state migrated over to the new data center, so that we can still have a tighter RPO when operations are resumed at the new site. 4. RELATED WORK Prior work on this topic falls into several categories: virtual machine migration, storage replication and network support. At the core of our technique is the ability of encapsulate applications within virtual machines that can be migrated without application downtimes [15]. Most virtual machine software, such as Xen P.11 [8] and VMWare [14] support live migration of VMs that involve extremely short downtimes ranging from tens of milliseconds to a second; details of Xen"s live migration techniques are discussed in [8]. As indicated earlier, these techniques assume that migration is being done on a LAN. VM migration has also been studied in the Shirako system [10] and for grid environments [17, 19]. Current virtual machine software support a suspend and resume feature that can be used to support WAN migration, but with downtimes [18, 12]. Recently live WAN migration using IP tunnels was demonstrated in [21], where an IP tunnel is set up from the source to destination server to transparently forward packets to and from the application; we advocate an alternate approach that assumes edge router support. In the context of storage, there exist numerous commercial products that perform replication, such as IBM Extended Remote Copy, HP Continuous Access XP, and EMC RepliStor. An excellent description of these and others, as well as a detailed taxonomy of the different approaches for replication can be found in [11]. The Ursa Minor system argues that no single fault model is optimal for all applications and proposed supporting data-type specific selections of fault models and encoding schemes for replication [1]. Recently, we proposed the notion of semantic-aware replication [13] where the system supports both synchronous and asynchronous replication concurrently and use signals from the file system to determine whether to replicate a particular write synchronously and asynchronously. In the context of network support, our work is related to the RouterFarm approach [2], which makes use of orchestrated network changes to realize near hitless maintenance on provider edge routers. In addition to being in a different application area, our approach differs from the RouterFarm work in two regards. First, we propose to have the required network changes be triggered by functionality outside of the network (as opposed to network management functions inside the network). Second, due to the stringent timing requirements of live migration, we expect that our approach would require new router functionality (as opposed to being realizable via the existing configuration interfaces). Finally, the recovery oriented computing (ROC) work emphasizes recovery from failures rather than failure avoidance [6]. In a similar spirit to ROC, we advocate using mechanisms from live VM migration to storage replication to support planned and unplanned outages in data centers (rather than full replication to mask such failures). 5. CONCLUSION P.12 A significant concern for Internet-based service providers is the continued operation and availability of services in the face of outages, whether planned or unplanned. In this paper we advocated a cooperative, context-aware approach to data center migration across WANs to deal with outages in a non-disruptive manner. We sought to achieve high availability of data center services in the face of both planned and incidental outages of data center facilities. We advocated using server virtualization technologies to enable the replication and migration of server functions. We proposed new network functions to enable server migration and replication across wide area networks (such as the Internet or a geographically distributed virtual private network), and finally showed the utility of intelligent and dynamic storage replication technology to ensure applications have access to data in the face of outages with very tight recovery point objectives. 6. REFERENCES [1] M. Abd-El-Malek, W. V. Courtright II, C. Cranor, G. R. Ganger, J. Hendricks, A. J. Klosterman, M. Mesnier, M. Prasad, B. Salmon, R. R. Sambasivan, S. Sinnamohideen, J. D. Strunk, E. Thereska, M. Wachs, and J. J. Wylie. Ursa minor: versatile cluster-based storage. USENIX Conference on File and Storage Technologies, December 2005. [2] Mukesh Agrawal, Susan Bailey, Albert Greenberg, Jorge Pastor, Panagiotis Sebos, Srinivasan Seshan, Kobus van der Merwe, and Jennifer Yates. Routerfarm: Towards a dynamic, manageable network edge. SIGCOMM Workshop on Internet Network Management (INM), September 2006. [3] L. Alvisi. Understanding the Message Logging Paradigm for Masking Process Crashes. PhD thesis, Cornell, January 1996. [4] L. Alvisi and K. Marzullo. Message logging: Pessimistic, optimistic, and causal. In Proceedings of the 15th International Conference on Distributed Computing Systems, pages 229-236. IEEE Computer Society, June 1995. 266 [5] Paul Barham, Boris Dragovic, Keir Fraser, Steven Hand, Tim Harris, Alex Ho, Rolf Neugebar, Ian Pratt, and Andrew Warfield. Xen and the art of virtualization. In the Proceedings of the ACM Symposium on Operating Systems Principles (SOSP), October 2003. [6] A. Brown and D. A. Patterson. Embracing failure: A case for recovery-oriented computing (roc). 2001 High Performance Transaction Processing Symposium, October 2001. [7] K. Brown, J. Katcher, R. Walters, and A. Watson. Snapmirror P.13 and snaprestore: Advances in snapshot technology. Network Appliance Technical Report TR3043. www. ne t app. c om/t e c h_ l i br ar y/3043. ht ml . [8] C. Clark, K. Fraser, S. Hand, J. Hanse, E. Jul, C. Limpach, I. Pratt, and A. Warfiel. Live migration of virtual machines. In Proceedings of NSDI, May 2005. [9] Disaster Recovery Journal. Business continuity glossary. ht t p: //www. dr j . c om/gl os s ar y/dr j gl os s ar y. ht ml . [10] Laura Grit, David Irwin, , Aydan Yumerefendi, and Jeff Chase. Virtual machine hosting for networked clusters: Building the foundations for autonomic orchestration. In In the First International Workshop on Virtualization Technology in Distributed Computing (VTDC), November 2006. [11] M. Ji, A. Veitch, and J. Wilkes. Seneca: Remote mirroring done write. USENIX 2003 Annual Technical Conference, June 2003. [12] M. Kozuch and M. Satyanarayanan. Internet suspend and resume. In Proceedings of the Fourth IEEE Workshop on Mobile Computing Systems and Applications, Calicoon, NY, June 2002. [13] Xiaotao Liu, Gal Niv, K. K. Ramakrishnan, Prashant Shenoy, and Jacobus Van der Merwe. The case for semantic aware remote replication. In Proc. 2nd International Workshop on Storage Security and Survivability (StorageSS 2006), Alexandria, VA, October 2006. [14] Michael Nelson, Beng-Hong Lim, and Greg Hutchins. Fast Transparent Migration for Virtual Machines. In USENIX Annual Technical Conference, 2005. [15] Mendel Rosenblum and Tal Garfinkel. Virtual machine monitors: Current technology and future trends. Computer, 38(5):39-47, 2005. [16] C. Ruemmler and J. Wilkes. Unix disk access patterns. Proceedings of Winter 1993 USENIX, Jan 1993. [17] Paul Ruth, Junghwan Rhee, Dongyan Xu, Rick Kennell, and Sebastien Goasguen. Autonomic Live Adaptation of Virtual Computational Environments in a Multi-Domain Infrastructure. In IEEE International Conference on Autonomic Computing (ICAC), June 2006. [18] Constantine P. Sapuntzakis, Ramesh Chandra, Ben Pfaff, Jim Chow, Monica S. Lam, and Mendel Rosenblum. Optimizing the migration of virtual computers. In Proceedings of the 5th Symposium on Operating Systems Design and P.14 Implementation, December 2002. [19] A. Sundararaj, A. Gupta, and P. Dinda. Increasing Application Performance in Virtual Environments through Run-time Inference and Adaptation. In Fourteenth International Symposium on High Performance Distributed Computing (HPDC), July 2005. [20] Symantec Corporation. Veritas Volume Replicator Administrator"s Guide. ht t p: //f t p. s uppor t . ve r i t as . c om/pub/s uppor t / pr oduc t s /Vol ume _ Re pl i c at or /2%83842. pdf , 5.0 edition, 2006. [21] F. Travostino, P. Daspit, L. Gommans, C. Jog, C. de Laat, J. Mambretti, I. Monga, B. van Oudenaarde, S. Raghunath, and P. Wang. Seamless live migration of virtual machines over the man/wan. Elsevier Future Generations Computer Systems, 2006. [22] T. Wood, P. Shenoy, A. Venkataramani, and M. Yousif. Black-box and gray-box strategies for virtual machine migration. In Proceedings of the Usenix Symposium on Networked System Design and Implementation (NSDI), Cambridge, MA, April 2007. [23] A xen way to iscsi virtualization? 2. Tập từ khóa xác định trước Tệp C-20.key thuộc kho ngữ liệu SemEval2010 gồm các từ khóa do tác giả xác định: internetbased, service, data, center, migration, wan, lan, virtual, server, storage, replication, synchronous, replication, asynchronous, replication, network, support, storage, voiceoverip, voip, database 3. Tập từ khóa ứng cử Kết quả trích rút tập từ khóa ứng cử lấy ra từ các Cụm danh từ, Thực thể có tên và các cụm ba từ xuất hiện lặp lại nhiều lần: migration, replication, virtualization, server, live, outages, center, storage, network, virtual, recovery, technologies, data, ramakrishnan, internetbased, application, remote, cooperative, wan, availability, maintenance, services, applications, service, technology, aware, distributed, operation, contextaware, management, approach, continued, centers, context, unplanned, access, providers, P.15 requirements, networks, significant, nondisruptive, business, systems, facilities, shenoy, prashant, intelligent, face, design, advances, objectives, disaster, manner, concern, categories, unanticipated, functions, wide, lan, high, minor, wans, propose, area, computercommunication, dynamic, internet, environment, mechanisms, physical, der, utility, disk, local, networking, functionality, failures, van, asynchronous, tight, ongoing, point, synchronous, jacobus, support, paper, disruptions, catastrophic, effective, current, servers, particular, critical, seamless, connectivity, administrator, tunnel, merwe, subsystems, environments, use, entertainment, operations, general, new, balance, labsresearch, outage, applicationservice, first, operating, second, available, address, users, essential, technical, location, allow, servicesapplications, realtime, building, events, continuous, instantaneous, introduction, milliseconds, sophisticated, reliability, semantics, descriptors, different, number, university, mirroring, machine, components, interactions, frequency, configuration, state, similar, private, system, techniques, abstract, downtime, amount, prior, concerns, underlying, connections, framework, work, changes, same, mission, ip, recent, robust, massachusetts, reasons, tool, contribution, coordinate, fashion, viewpoint, reachability, multiple, level, terms, redundancy, subject, example, failure, presents, challenges, performance, blocks, requirement, disruption, provider, section, http, software, platforms, extensions, write, networkbased, hotswappable, active, traffic, case, redundant, kk, such, today, throughput, robustness, running, shared, practices, knowledge, little, essence, appropriate, capabilities, businessusability, large, individual, needs, tens, consistent, es, addresses, unsolicited, cognizant, implication, necessity, complete, supplies, ability, load, router, nature, process, towards, databases, eg, primary, coordination, becomes, mobility, replicas, unique, experience, checkpoint, optimal, devices, efficiency, desirable, entirety, varies, attractive, difficulty, unresolved, anticipation, entire, san, switchover, observations, convergence, feat, hundreds, logic, constraints, applicationaware, way, main, simultaneous, means, unavailability, localprimary, reason, coalescing, challenge, other, considerations, actions, part, driver, initial, consistency, binding, latency, overhead, extension, whole, desire, weight, switch, latter, preparation, signal, snapshot, anticipated, actual, subset, divergence, space, operators, transfer, localized, enabler, alternative, processor, sessionbased, heavy, protocols, layertwo, discussion, power, phase, vpns, decades, focus, purposes, kind, routers, completion, loss, basic, preestablished, routed, event, site, signals, properties, P.16 consequence, key, recoveryoriented, financial, edge, necessary, communication, impact, perspective, intermediate, virtualized, subsystem, order, convenient, time, previous, switches, routing, orchestrated, reply, approaches, checkpoints, cases, suspend, bulk, information, encapsulate, preferred, modern, efficient, solution, mac, priority, concert, nonlan, manual, years, clusterbased, sans, returns, flexible, need, scales, semanticaware, problem, secondary, occurrence, path, initiation, subsequent, builds, writes, direct, status, metrics, starts, forms, change, situation, exposition, addition, selections, scenarios, resume, gateway, issues, disks, voiceoverip, persistent, affordable, certain, delivery, exposure, copies, successful, block, detail, simplicity, ie, orchestration, snapshots, situations, description, incidental, final, emphasizes, sessions, normal, element, proceeds, checkpointing, effects, customer, periodic, route, commercial, stringent, interfaces, extreme, spectrum, instance, over, blades, principle, processing, foundations, avoidance, arp, scale, manageable, mechanism, datatype, alternate, standbyin, continuity, excellent, snaprestore, technique, versatile, mode, step, downtimes, price, ingress, minimum, fact, prompting, replica, impending, activity, session, taxonomy, advanced, memory, interval, products, numerous, parts, safe, control, view, clients, standby, function, detailed, figure, steps, regards, strategies, dormant, specific, deeper, computers, vs, proxy, monitors, fault, clusters, basis, right, machines, course, details, feature, points, schemes, quick, cold, voip, differs, timing, semantic, incremental, several, tunnels, graybox, patterns, full, hitless, ht, glossary, spirit, causal, international, notion, source, usenix, set, models, others, edition, thesis, topic, ar, single, note, hot, survivability, short, proceedings, trends, autonomic, implementation, pe, grid, warm, core, infrastructure, venkataramani, model, copy, computing, many, 15th, symposium, ml, ce, pages, manwan, computational, routerfarm, conference, y3043, multidomain, transaction, www, os, uppor, transparent, generations, networked, conclusion, ne, december, gl, replicator, art, pdf, adaptation, principles, logging, roc, pes, replistor, 5th, fourteenth, abdelmalek, garfinkel, workshop, dr, related, sebastien, ydr, references, computer, pl, pr, paradigm, crashes, sundararaj, inference, agrawal, appliance, travostino, oudenaarde, cornell, sigcomm, thereska, message, blackbox, prasad, seshan, rpos, ve, security, dragovic, vm, raghunath, society, file, calicoon, mesnier, benghong, june, goasguen, nelson, journal, runtime, winter, annual, cranor, sip, pastor, dongyan, barham, fourth, ieee, report, wilkes, tr3043, future, de, xp, sosp, seneca, volume, unix, wachs, gupta, icac, guide, yousif, wood, gal, mms, phd, dinda, proc, P.17 lim, vtdc, wang, liu, jan, inm, rhee, andrew, hpdc, ii, va, sinnamohideen, niv, monica, xu, tal, ny, satyanarayanan, hp, kobus, re, keir, alexandria, srinivasan, rosenblum, hendricks, storagess, laat, marzullo, courtright, shirako, klosterman, sambasivan, panagiotis, vmware, yumerefendi, corporation, snapmirror, xen, greenberg, bailey, hutchins, constantine, sapuntzakis, patterson, albert, app, harris, jennifer, nsdi, steven, neugebar, warfield, ursa, watson, ibm, paul, susan, chandra, kozuch, brown, vms, ganger, salmon, strunk, katcher, walters, irwin, clark, mukesh, alvisi, limpach, warfiel, emc, fraser, alex, michael, ruemmler, mambretti, chase, junghwan, ramesh, wylie, xiaotao, jorge, sebos, yates, symantec, pratt, cambridge, boris, jeff, tim, kennell, ian, david, oduc, veitch, hand, ho, rick, mendel, hanse, veritas, gommans, laura, aydan, rolf, daspit, grit, acm, lam, ben, jim, pfaff, jul, greg, ji, monga, ruth, chow, vol, jog, ume, ma, rpo 4. Kết quả xác định độ quan trọng của từ Kết quả dự đoán độ quan trọng của mỗi từ trong tập từ khóa ứng cử và sắp xếp giảm dần theo độ quan trọng (Các từ in đậm trong 10 kết quả đầu tiên nằm trong tập từ khóa xác định trước). Bảng P1. Kết quả dự đoán độ quan trọng của từ STT Từ khóa ứng cử Giá trị y STT Từ khóa ứng cử Giá trị y 1 migration 0.840984 382 resume 0.0045808 2 replication 0.678164 383 gateway 0.0045794 3 virtualization 0.659803 384 issues 0.004577 4 server 0.618527 385 disks 0.0045655 5 live 0.61789 386 voiceoverip 0.0045531 6 outages 0.610963 387 persistent 0.0045473 7 center 0.606313 388 affordable 0.004527 8 storage 0.542996 389 certain 0.0045241 9 network 0.533428 390 delivery 0.0045182 10 virtual 0.490329 391 exposure 0.0045027 11 recovery 0.462664 392 copies 0.0044478 12 technologies 0.439565 393 successful 0.0044444 13 data 0.439148 394 block 0.0044061 14 ramakrishnan 0.371115 395 detail 0.0043981 15 internetbased 0.363683 396 simplicity 0.0043912 16 application 0.32836 397 ie 0.0043785 17 remote 0.311041 398 orchestration 0.0043709 18 cooperative 0.298606 399 snapshots 0.00437 19 wan 0.260697 400 situations 0.0043109 20 availability 0.249949 401 description 0.0042526 21 maintenance 0.246622 402 incidental 0.0042457 P.18 STT Từ khóa ứng cử Giá trị y STT Từ khóa ứng cử Giá trị y 22 services 0.236617 403 final 0.0042226 23 applications 0.21758 404 emphasizes 0.0042183 24 service 0.208958 405 sessions 0.0041841 25 technology 0.188583 406 normal 0.0041711 26 aware 0.184637 407 element 0.0041647 27 distributed 0.184114 408 proceeds 0.0041553 28 operation 0.178693 409 checkpointing 0.0041511 29 contextaware 0.174755 410 effects 0.0041382 30 management 0.16334 411 customer 0.00413 31 approach 0.160562 412 periodic 0.0040518 32 continued 0.159632 413 route 0.0040377 33 centers 0.151135 414 commercial 0.0040179 34 context 0.149825 415 stringent 0.0039577 35 unplanned 0.136594 416 interfaces 0.0039557 36 access 0.135795 417 extreme 0.0039349 37 providers 0.132818 418 spectrum 0.0039304 38 requirements 0.132789 419 instance 0.0039263 39 networks 0.131715 420 over 0.003924 40 significant 0.129641 421 blades 0.0039219 41 nondisruptive 0.127865 422 principle 0.0038941 42 business 0.126391 423 processing 0.0038913 43 systems 0.123377 424 foundations 0.0038743 44 facilities 0.122692 425 avoidance 0.0038511 45 shenoy 0.113144 426 arp 0.0038416 46 prashant 0.109356 427 scale 0.0038255 47 intelligent 0.107079 428 manageable 0.0038238 48 face 0.10703 429 mechanism 0.0038141 49 design 0.10576 430 datatype 0.0037792 50 advances 0.105361 431 alternate 0.0037481 51 objectives 0.103294 432 standbyin 0.0037446 52 disaster 0.102727 433 continuity 0.0037397 53 manner 0.102325 434 excellent 0.0037301 54 concern 0.101271 435 snaprestore 0.0036539 55 categories 0.098923 436 technique 0.0036314 56 unanticipated 0.097049 437 versatile 0.0036119 57 functions 0.097024 438 mode 0.0036081 58 wide 0.094731 439 step 0.0036024 59 lan 0.09354 440 downtimes 0.0035835 60 high 0.091383 441 price 0.0035753 61 minor 0.087343 442 ingress 0.0035746 62 wans 0.081532 443 minimum 0.0035724 63 propose 0.081235 444 fact 0.0035698 P.19 STT Từ khóa ứng cử Giá trị y STT Từ khóa ứng cử Giá trị y 64 area 0.075401 445 prompting 0.0035426 65 computercommunication 0.071124 446 replica 0.0035229 66 dynamic 0.070823 447 impending 0.0035184 67 internet 0.070188 448 activity 0.003507 68 environment 0.07006 449 session 0.0034768 69 mechanisms 0.068699 450 taxonomy 0.0034586 70 physical 0.067211 451 advanced 0.0033992 71 der 0.065018 452 memory 0.0033911 72 utility 0.064928 453 interval 0.0033575 73 disk 0.063486 454 products 0.0033335 74 local 0.062574 455 numerous 0.0032772 75 networking 0.062066 456 parts 0.0032592 76 functionality 0.060316 457 safe 0.0032163 77 failures 0.058608 458 control 0.0031983 78 van 0.056861 459 view 0.0031885 79 asynchronous 0.055627 460 clients 0.0031725 80 tight 0.055299 461 standby 0.0031548 81 ongoing 0.055233 462 function 0.0031449 82 point 0.055094 463 detailed 0.0031171 83 synchronous 0.055003 464 figure 0.0030952 84 jacobus 0.054747 465 steps 0.0030943 85 support 0.05382 466 regards 0.0030621 86 paper 0.053773 467 strategies 0.0030361 87 disruptions 0.053526 468 dormant 0.0029639 88 catastrophic 0.052759 469 specific 0.0029358 89 effective 0.052542 470 deeper 0.0029295 90 current 0.051912 471 computers 0.0029269 91 servers 0.049387 472 vs 0.0029162 92 particular 0.049132 473 proxy 0.0029011 93 critical 0.047906 474 monitors 0.0028663 94 seamless 0.046336 475 fault 0.00285 95 connectivity 0.044587 476 clusters 0.002828 96 administrator 0.043383 477 basis 0.002826 97 tunnel 0.04324 478 right 0.0028248 98 merwe 0.041538 479 machines 0.0028083 99 subsystems 0.041409 480 course 0.0028024 100 environments 0.040257 481 details 0.0027973 101 use 0.036387 482 feature 0.0027855 102 entertainment 0.036088 483 points 0.0027779 103 operations 0.03538 484 schemes 0.0027776 104 general 0.034748 485 quick 0.0027557 105 new 0.034575 486 cold 0.0027474 P.20 STT Từ khóa ứng cử Giá trị y STT Từ khóa ứng cử Giá trị y 106 balance 0.034293 487 voip 0.0027359 107 labsresearch 0.034123 488 differs 0.0027316 108 outage 0.032523 489 timing 0.0026993 109 applicationservice 0.0323 490 semantic 0.0026863 110 first 0.032053 491 incremental 0.0026836 111 operating 0.031873 492 several 0.002665 112 second 0.031871 493 tunnels 0.0026577 113 available 0.031776 494 graybox 0.0026532 114 address 0.031016 495 patterns 0.0025977 115 users 0.030828 496 full 0.0025534 116 essential 0.030411 497 hitless 0.0025488 117 technical 0.030131 498 ht 0.0025404 118 location 0.029677 499 glossary 0.0025208 119 allow 0.029313 500 spirit 0.002509 120 servicesapplications 0.029243 501 causal 0.0024637 121 realtime 0.029228 502 international 0.002438 122 building 0.029168 503 notion 0.0024247 123 events 0.028764 504 source 0.0024038 124 continuous 0.028121 505 usenix 0.0023922 125 instantaneous 0.027937 506 set 0.0023618 126 introduction 0.027788 507 models 0.0023371 127 milliseconds 0.027768 508 others 0.0023345 128 sophisticated 0.027693 509 edition 0.0022988 129 reliability 0.027484 510 thesis 0.0022935 130 semantics 0.02676 511 topic 0.0022763 131 descriptors 0.025592 512 ar 0.0022534 132 different 0.025459 513 single 0.0022509 133 number 0.025158 514 note 0.0022465 134 university 0.024702 515 hot 0.0022409 135 mirroring 0.024501 516 survivability 0.0022022 136 machine 0.024371 517 short 0.0021964 137 components 0.024161 518 proceedings 0.0021695 138 interactions 0.024085 519 trends 0.0021362 139 frequency 0.023768 520 autonomic 0.0021327 140 configuration 0.023489 521 implementation 0.002083 141 state 0.023362 522 pe 0.0020795 142 similar 0.023309 523 grid 0.0020748 143 private 0.023003 524 warm 0.002055 144 system 0.022504 525 core 0.0020492 145 techniques 0.021737 526 infrastructure 0.0020421 146 abstract 0.021286 527 venkataramani 0.0020299 147 downtime 0.021142 528 model 0.0020187 P.21 STT Từ khóa ứng cử Giá trị y STT Từ khóa ứng cử Giá trị y 148 amount 0.020983 529 copy 0.00194 149 prior 0.020438 530 computing 0.0018993 150 concerns 0.020368 531 many 0.0018784 151 underlying 0.020181 532 15th 0.0018723 152 connections 0.020169 533 symposium 0.0018621 153 framework 0.019967 534 ml 0.0018518 154 work 0.019943 535 ce 0.0018451 155 changes 0.019746 536 pages 0.0017967 156 same 0.019699 537 manwan 0.0017894 157 mission 0.019429 538 computational 0.0017754 158 ip 0.019258 539 routerfarm 0.0017595 159 recent 0.018481 540 conference 0.0017545 160 robust 0.018437 541 y3043 0.00175 161 massachusetts 0.018398 542 multidomain 0.0017365 162 reasons 0.018173 543 transaction 0.0017094 163 tool 0.018069 544 www 0.0016864 164 contribution 0.017874 545 os 0.0016773 165 coordinate 0.017508 546 uppor 0.0016162 166 fashion 0.01731 547 transparent 0.0016075 167 viewpoint 0.017007 548 generations 0.0015975 168 reachability 0.016715 549 networked 0.0015931 169 multiple 0.016661 550 conclusion 0.0015557 170 level 0.016497 551 ne 0.0015469 171 terms 0.016478 552 december 0.0015402 172 redundancy 0.016408 553 gl 0.0015227 173 subject 0.015692 554 replicator 0.0015208 174 example 0.015675 555 art 0.0015121 175 failure 0.015571 556 pdf 0.0015027 176 presents 0.015501 557 adaptation 0.001501 177 challenges 0.015161 558 principles 0.0014718 178 performance 0.014878 559 logging 0.0014657 179 blocks 0.014839 560 roc 0.0014649 180 requirement 0.014562 561 pes 0.0014648 181 disruption 0.014552 562 replistor 0.0014273 182 provider 0.014527 563 5th 0.0014269 183 section 0.014516 564 fourteenth 0.0014263 184 http 0.014488 565 abdelmalek 0.0014201 185 software 0.01448 566 garfinkel 0.0014167 186 platforms 0.014328 567 workshop 0.0014111 187 extensions 0.014232 568 dr 0.0013682 188 write 0.014197 569 related 0.0013521 189 networkbased 0.014171 570 sebastien 0.0013487 P.22 STT Từ khóa ứng cử Giá trị y STT Từ khóa ứng cử Giá trị y 190 hotswappable 0.013964 571 ydr 0.0013394 191 active 0.01382 572 references 0.0013047 192 traffic 0.013746 573 computer 0.0012926 193 case 0.013394 574 pl 0.0012628 194 redundant 0.013225 575 pr 0.0012412 195 kk 0.012854 576 paradigm 0.0012372 196 such 0.01283 577 crashes 0.0012056 197 today 0.01281 578 sundararaj 0.001189 198 throughput 0.012766 579 inference 0.001188 199 robustness 0.012753 580 agrawal 0.0011803 200 running 0.012621 581 appliance 0.001175 201 shared 0.012589 582 travostino 0.0011587 202 practices 0.012569 583 oudenaarde 0.0011525 203 knowledge 0.012438 584 cornell 0.0011472 204 little 0.012241 585 sigcomm 0.00111 205 essence 0.012232 586 thereska 0.0011092 206 appropriate 0.012105 587 message 0.0011037 207 capabilities 0.012089 588 blackbox 0.0011029 208 businessusability 0.011849 589 prasad 0.0010992 209 large 0.011811 590 seshan 0.0010836 210 individual 0.011753 591 rpos 0.0010718 211 needs 0.011703 592 ve 0.0010707 212 tens 0.01148 593 security 0.0010704 213 consistent 0.011461 594 dragovic 0.0010703 214 es 0.011417 595 vm 0.001057 215 addresses 0.01139 596 raghunath 0.0010229 216 unsolicited 0.011261 597 society 0.001008 217 cognizant 0.011184 598 file 0.0010009 218 implication 0.011171 599 calicoon 0.0009904 219 necessity 0.011126 600 mesnier 0.0009895 220 complete 0.011042 601 benghong 0.000976 221 supplies 0.010953 602 june 0.0009648 222 ability 0.010771 603 goasguen 0.0009563 223 load 0.010654 604 nelson 0.0009411 224 router 0.010647 605 journal 0.0009258 225 nature 0.010615 606 runtime 0.000911 226 process 0.010454 607 winter 0.0009097 227 towards 0.009911 608 annual 0.0008994 228 databases 0.009863 609 cranor 0.0008817 229 eg 0.009755 610 sip 0.0008663 230 primary 0.009714 611 pastor 0.0008657 231 coordination 0.009567 612 dongyan 0.000851 P.23 STT Từ khóa ứng cử Giá trị y STT Từ khóa ứng cử Giá trị y 232 becomes 0.009501 613 barham 0.0008447 233 mobility 0.009493 614 fourth 0.0008392 234 replicas 0.009441 615 ieee 0.0008385 235 unique 0.009179 616 report 0.000827 236 experience 0.009144 617 wilkes 0.0008237 237 checkpoint 0.009134 618 tr3043 0.0008223 238 optimal 0.009079 619 future 0.0008174 239 devices 0.009068 620 de 0.0008126 240 efficiency 0.009034 621 xp 0.0007982 241 desirable 0.009008 622 sosp 0.0007951 242 entirety 0.008996 623 seneca 0.0007899 243 varies 0.008913 624 volume 0.0007782 244 attractive 0.008811 625 unix 0.0007777 245 difficulty 0.008741 626 wachs 0.0007764 246 unresolved 0.008637 627 gupta 0.0007693 247 anticipation 0.008615 628 icac 0.0007677 248 entire 0.008518 629 guide 0.0007181 249 san 0.008446 630 yousif 0.0007099 250 switchover 0.008434 631 wood 0.0006844 251 observations 0.008378 632 gal 0.0006833 252 convergence 0.008318 633 mms 0.0006742 253 feat 0.007965 634 phd 0.0006715 254 hundreds 0.007892 635 dinda 0.0006558 255 logic 0.00782 636 proc 0.0006469 256 constraints 0.007783 637 lim 0.0006428 257 applicationaware 0.007772 638 vtdc 0.0006254 258 way 0.00776 639 wang 0.0006206 259 main 0.007727 640 liu 0.000605 260 simultaneous 0.007708 641 jan 0.0006041 261 means 0.007664 642 inm 0.000602 262 unavailability 0.007623 643 rhee 0.0005962 263 localprimary 0.007621 644 andrew 0.0005909 264 reason 0.007501 645 hpdc 0.0005783 265 coalescing 0.00743 646 ii 0.0005641 266 challenge 0.007417 647 va 0.0005618 267 other 0.007394 648 sinnamohideen 0.0005494 268 considerations 0.00738 649 niv 0.0005456 269 actions 0.007338 650 monica 0.0005414 270 part 0.007235 651 xu 0.0005404 271 driver 0.007214 652 tal 0.0005352 272 initial 0.007191 653 ny 0.0005255 273 consistency 0.007187 654 satyanarayanan 0.0004802 P.24 STT Từ khóa ứng cử Giá trị y STT Từ khóa ứng cử Giá trị y 274 binding 0.007176 655 hp 0.0004593 275 latency 0.007136 656 kobus 0.0004488 276 overhead 0.007092 657 re 0.0004384 277 extension 0.007062 658 keir 0.0003816 278 whole 0.007004 659 alexandria 0.0003689 279 desire 0.006951 660 srinivasan 0.0003569 280 weight 0.006927 661 rosenblum 0.0003449 281 switch 0.006921 662 hendricks 0.0003437 282 latter 0.006908 663 storagess 0.0003361 283 preparation 0.006881 664 laat 0.0002986 284 signal 0.006875 665 marzullo 0.0002801 285 snapshot 0.006874 666 courtright 0.0002798 286 anticipated 0.006833 667 shirako 0.0002795 287 actual 0.006811 668 klosterman 0.0002775 288 subset 0.006804 669 sambasivan 0.0002757 289 divergence 0.006801 670 panagiotis 0.0002679 290 space 0.006789 671 vmware 0.0002562 291 operators 0.006741 672 yumerefendi 0.0002521 292 transfer 0.006687 673 corporation 0.0002517 293 localized 0.006684 674 snapmirror 0.0002388 294 enabler 0.006667 675 xen 0.0002384 295 alternative 0.006654 676 greenberg 0.0002366 296 processor 0.00665 677 bailey 0.0002238 297 sessionbased 0.006645 678 hutchins 0.000217 298 heavy 0.00664 679 constantine 0.0002148 299 protocols 0.006616 680 sapuntzakis 0.0002146 300 layertwo 0.006604 681 patterson 0.000214 301 discussion 0.006598 682 albert 0.0002124 302 power 0.006525 683 app 0.0002115 303 phase 0.006509 684 harris 0.0002076 304 vpns 0.006507 685 jennifer 0.0002048 305 decades 0.006464 686 nsdi 0.0001985 306 focus 0.006448 687 steven 0.0001931 307 purposes 0.006413 688 neugebar 0.0001924 308 kind 0.006337 689 warfield 0.0001916 309 routers 0.006307 690 ursa 0.0001906 310 completion 0.006291 691 watson 0.0001886 311 loss 0.006257 692 ibm 0.0001831 312 basic 0.00625 693 paul 0.0001793 313 preestablished 0.006211 694 susan 0.0001779 314 routed 0.006206 695 chandra 0.0001771 315 event 0.006204 696 kozuch 0.0001758 P.25 STT Từ khóa ứng cử Giá trị y STT Từ khóa ứng cử Giá trị y 316 site 0.006195 697 brown 0.0001757 317 signals 0.006155 698 vms 0.0001715 318 properties 0.006106 699 ganger 0.000165 319 consequence 0.006096 700 salmon 0.0001633 320 key 0.006056 701 strunk 0.0001625 321 recoveryoriented 0.006006 702 katcher 0.0001613 322 financial 0.005989 703 walters 0.000161 323 edge 0.00598 704 irwin 0.0001609 324 necessary 0.00597 705 clark 0.0001604 325 communication 0.005953 706 mukesh 0.0001593 326 impact 0.005874 707 alvisi 0.0001563 327 perspective 0.005872 708 limpach 0.0001553 328 intermediate 0.005856 709 warfiel 0.0001546 329 virtualized 0.005826 710 emc 0.0001529 330 subsystem 0.00579 711 fraser 0.0001527 331 order 0.005773 712 alex 0.0001513 332 convenient 0.005768 713 michael 0.0001502 333 time 0.005661 714 ruemmler 0.0001499 334 previous 0.005661 715 mambretti 0.0001485 335 switches 0.005652 716 chase 0.0001482 336 routing 0.005564 717 junghwan 0.0001478 337 orchestrated 0.00556 718 ramesh 0.0001438 338 reply 0.00556 719 wylie 0.0001412 339 approaches 0.005505 720 xiaotao 0.0001395 340 checkpoints 0.005482 721 jorge 0.0001386 341 cases 0.005471 722 sebos 0.0001385 342 suspend 0.005456 723 yates 0.0001374 343 bulk 0.005414 724 symantec 0.0001365 344 information 0.005397 725 pratt 0.0001327 345 encapsulate 0.005384 726 cambridge 0.0001317 346 preferred 0.005355 727 boris 0.0001299 347 modern 0.005348 728 jeff 0.0001295 348 efficient 0.005325 729 tim 0.0001289 349 solution 0.005286 730 kennell 0.0001287 350 mac 0.005283 731 ian 0.0001283 351 priority 0.005279 732 david 0.0001278 352 concert 0.005264 733 oduc 0.0001276 353 nonlan 0.005264 734 veitch 0.0001266 354 manual 0.005255 735 hand 0.000125 355 years 0.005195 736 ho 0.0001217 356 clusterbased 0.005181 737 rick 0.0001214 357 sans 0.005139 738 mendel 0.0001195 P.26 STT Từ khóa ứng cử Giá trị y STT Từ khóa ứng cử Giá trị y 358 returns 0.005111 739 hanse 0.0001193 359 flexible 0.005098 740 veritas 0.0001192 360 need 0.005092 741 gommans 0.0001144 361 scales 0.005086 742 laura 0.0001142 362 semanticaware 0.005044 743 aydan 0.000114 363 problem 0.004982 744 rolf 0.0001129 364 secondary 0.004961 745 daspit 0.0001003 365 occurrence 0.004953 746 grit 0.0001001 366 path 0.004924 747 acm 9.91E-05 367 initiation 0.004921 748 lam 9.84E-05 368 subsequent 0.004894 749 ben 9.84E-05 369 builds 0.004858 750 jim 9.81E-05 370 writes 0.004834 751 pfaff 9.61E-05 371 direct 0.004823 752 jul 9.06E-05 372 status 0.004808 753 greg 9.00E-05 373 metrics 0.004762 754 ji 8.76E-05 374 starts 0.004734 755 monga 8.70E-05 375 forms 0.004701 756 ruth 8.66E-05 376 change 0.004692 757 chow 8.39E-05 377 situation 0.004689 758 vol 6.78E-05 378 exposition 0.004662 759 jog 6.70E-05 379 addition 0.00465 760 ume 5.71E-05 380 selections 0.004642 761 ma 5.24E-05 381 scenarios 0.004629 762 rpo 1.36E-05