Peter William Dickman. Distributed Object Management in a Non-Small Graph of Autonomous Networks with Few Failures. PhD thesis, University of Cambridge, september 1991. 7  Home/Search   Document Not in Database   Summary   Related Articles   Check

....problems like load sharing, fault tolerance and efficient resources access. Numerous studies have been completed on this problems: Utopia [3] Charlotte [4] and Gatos [5] for load balancing, Coda [6] for disconnected file operations, Amoeba [7] Star[8] for fault tolerance. and object management [9]. This management is even more complex in the case of mobile computers [10] due to: ffl frequent disconnection of mobile computers, ffl resources duplication to allow disconnect work, ffl reconnection after a location s change, 2.1 Territory Management Scheme In our system, an execution ....

Peter William Dickman. Distributed Object Management in a Non-Small Graph of Autonomous Networks with Few Failures. PhD thesis, University of Cambridge, september 1991. 7

....a cyclic one. Usually, the cyclic distributed garbage collector is triggered at a low rate and most of garbage is assumed to be reclaimed by the acyclic one. However, to be efficient such combination relies on the assumption that global tracing frequency is low compared to the acyclic collector. Dickman [1991] combines his Optimised Weighted Reference Counting with an (unspecified) cyclic global garbage collector. This cyclic collector is responsible for both reclaiming distributed cycles and objects Optimised Weighted Reference Counting can t collect (explained in Section 3.4) Triggering the cyclic ....

....group composed of spaces A and B. Table 1: Taxonomy of some popular distributed GC techniques. main distributed GC characteristics Distributed GC cyclic floating large imple space message failure Techniques garbage scale mented failure duplic. loss late Reference Counting [Bevan, 1987] ffl [Dickman, 1991] ffi ffl ffl ffl [Piquer, 1991] ffl ffl ffl ffl [Goldberg, 1989] ffl Reference Listing [Shapiro et al. 1990] ffl ffl ffl ffl ffl ffl ffl [Birrell et al. 1993] ffl ffl ffl ffi ffl ffl ffl Tracing [Ali, 1984] ffl [Hughes, 1985] ffl ffl [Ladin and Liskov, 1992] ffl ffl [Lang et al. 1992] ....

Peter Dickman. Distributed Object Management in a Non-Small Graph of Autonomous Networks With Few Failures. PhD thesis, University of Cambridge, United Kingdom, September 1991.

....and does not dictate the nature of the GGD algorithm. However, the choice of strategy used by the GGD affects the nature and amount of information the log keeping mechanism has to maintain. For instance, GGD approaches based on weighted reference counting [Bevan, 1987, Watson and Watson, 1987, Dickman, 1991] or reference listing [Plainfoss e, 1994b] makes it possible to avoid eager logkeeping but are not intrinsically comprehensive. This document describes a lazy log keeping facility aimed at supporting comprehensive GGD on Amadeus [Cahill et al. 1993] 2 System Model This section presents an ....

Dickman, P. W. (1991). Distributed Object Management in a Non-Small Graph of Autonomous Networks with Few Failures. PhD thesis, Darwin College, Cambridge University.

.... of Amadeus [2, 11] a general purpose fine grained object oriented distributed system, supporting applications written in several languages, we present a general and scalable dynamic clustering algorithm, possibly programmerdriven, tightly coupled to garbage collection and load balancing 1 [4, 14], and executing in parallel with user computations. 1 In Amadeus, load balancing can be configured for balancing activities (distributed threads of control) or object clusters. This paper is structured as follows. First, an introduction to Amadeus, including the actual clustering policy, is ....

Peter W. Dickman. Distributed Object Management in a Non-Small Gra ph of Autonomous Networks with Few Failures. PhD thesis, University of Cambridge Computer Laboratory, 1992.

....systems, calling systems where this unit is small (i.e. a few hundred bytes of data and a few thousand bytes of code) fine grained, while systems where this unit is much larger (i.e. several kilobytes of data and code) are called coarse grained. Examples of fine grained systems are Bellerophon [DIC 92a] Ellie [AND 92] and Emerald [JUL 87] while other systems like Eden [ALM 85] Sprite [DOU 91] and UNIX [ZHO 92] are regarded as coarse grained. The load distribution facility may use the unit of distribution directly or aggregate these units to build a larger unit of load distribution (e.g. ....

Peter Dickman: "Distributed Object Management in a Non-Small Graph of Autonomous networks with Few Failures", Ph.D. Thesis, University of Cambridge Computer Laboratory, Cambridge, England, 1992.

....Otherwise live objects could erroneously be identified as garbage. This consistency constraint can therefore potentially be both costly (in terms of additional messages for instance) and complex when eager log keeping is chosen. GGD approaches based on weighted reference counting [Bev87, WW87, Dic91] or reference listing [Pla94b] makes it possible to avoid this form of eager log keeping but are not intrinsically comprehensive. This document describes a lazy log keeping facility aimed at supporting comprehensive GGD on Amadeus [CBSH93] 2 System Model This section presents an abstract view ....

Peter William Dickman. Distributed Object Management in a Non-Small Graph of Autonomous Networks with Few Failures. PhD thesis, Darwin College, Cambridge University, September 1991.

....to maintaining additional information depending on the choice of Site 1 Site 2 A B C D F E G E F A C root 1 root 1 root 2 Figure 2: An object graph and its global root graph. GGD strategy. For instance the contents of the logs may consist of a weight as in weighted reference counting schemes [2, 19, 6] or the identity of the recipient of a reference as in reference listing schemes [15] These logs, which may be either centralised or distributed, together constitute a consistent, although not necessarily complete, snapshot of the object graph. The logs are consistent if they reflect a ....

.... approaches are therefore generally believed to be necessarily unscalable [15] As a consequence, comprehensiveness has often been tradedoff for scalability under the assumptions that distributed cycles are relatively rare, and that only graph tracing algorithms can be intrinsically comprehensive [6, 2, 19]. Instead, in the absence of empirical evidence to the contrary, we contend that distributed cycles of garbage are as likely to occur as local cycles, and that intrinsically comprehensive GGD algorithms can, in fact, be scalable as well. Our alternative to graph tracing consists in analyzing the ....

P. W. Dickman. Distributed Object Management in a Non-Small Graph of Autonomous Networks with Few Failures. PhD thesis, Darwin College, Cambridge University (UK), Sep. 91.

....3.1, since it is not an issue when the communication protocol is atomic, and a causal protocol makes it simpler. The two main points on this axis are the standard GC algorithms, counting and tracing. Most published algorithms combine counting and tracing to some degree. For instance in Dickman [7] most, but not all, unreachable scions are detected by counting; the remainder are detected by a complementary global trace. Counting variants are either non fault tolerant [1, 21, 29] or fault tolerant [26] see related Sections 3.1, 3.2) The former are of course simpler than the latter. They ....

....a complementary global trace. Counting variants are either non fault tolerant [1, 21, 29] or fault tolerant [26] see related Sections 3.1, 3.2) The former are of course simpler than the latter. They also differ on how they deal with cycles of garbage: by migration [4] by a complementary trace [7], or not at all [1, 21, 29, etc. Tracing is feasible only in a small scale system. Variants within the tracing family include global tracing [16, 28] centralized tracing [17] tracing in groups [18] and timestamped asynchronous global tracing [14] Although all require some global ....

[Article contains additional citation context not shown here]

Peter William Dickman. Distributed Object Management in a Non-Small Graph of Autonomous Networks with Few Failures. PhD thesis, Darwin College, U. of Cambridge, Cambridge, England (GB), March 1992.

.... the constraints imposed by distribution, is to trade off comprehensiveness, i.e. the ability to detect distributed cycles of garbage, for weaker inter node synchronisation constraints and a higher degree of concurrency under the assumption that distributed cycles are, in fact, relatively rare [8, 7, 2, 16]. Under this assumption, it can be considered acceptable for instance, to try to detect these rare cycles [13] by heuristically co locating objects likely to be part of a cycle so that they can be dealt with by some local comprehensive GC algorithm a la Bishop [3] Instead, we prefer to make no ....

....must be avoided. Otherwise live objects could erroneously be identified as garbage. This consistency constraint can therefore potentially be both costly (in terms of additional messages for instance) and complex when eager logkeeping is chosen. GGD approaches based on weighted reference counting [2, 16, 7] or reference listing [13] makes it possible to avoid this form of eager logkeeping but are not intrinsically comprehensive. 5 Consensus free GGD Alternative Schelvis proposed a GGD algorithm based on the asynchronous and incremental distribution of timestamp packets [14] which seems to have been ....

Peter William Dickman. Distributed Object Management in a Non-Small Graph of Autonomous Networks with Few Failures. PhD thesis, Darwin College, Cambridge University, September 1991.

....they only become garbage when other processes of the distributed application have finished as well. Trying to prevent the storage of objects that are garbage at the termination of a given process requires an on line algorithm. Due to the distributed cycles, refence counting algorithms (e.g. [8]) are not very useful here. The same applies to protocols tightening incoming and outcomming references [9] We could follow Lieberman suggestion and evacuate objects from termi 3 COLLECTING PERSISTENT OBJECTS 4 nating processes to remaining ones that still refer to them [10] However, this is ....

Peter W. Dickman. Distributed Object Management in a Non-Small Graph of Autonomous Networks with Few Failures. PhD thesis, University of Cambridge Computer Laboratory, 1992.

....garbage. Both reference listing and reference counting schemes require that cyclic garbage be rare and sufficient memory be provided to tolerate the leakage. Extensions to reference listing to handle cycles include optimised weighted reference counting augmented with background global tracing [Dickman91], and reference listing with partial tracing [RJ96] 24 8.3 Tracing Hughes [Hughes85] uses time stamps based on global time to trace live objects. Each trace initiated on a node uses the time stamp to mark objects. Each outgoing pointer uses the time stamp whenever it propagates the trace to ....

Peter Dickman. Distributed object management in a non-small graph of autonomous networks with few failures. PhD thesis, University of Cambridge, United Kingdom, September 1991.

....above, by relying on the existence of a global garbage collector (which is necessary anyway to collect distributed cycles of garbage, since they are not removed by the protocol presented in this paper) to detect and remove garbage scions retained in this way. This is the approach taken by Dickman [6] and means that the algorithm is no longer live, but remains efficient and effective. An alternative would make the rule for broken chains be the same as for objects: any SSP chain indirecting through a terminated space would be deemed dangling. Care must be taken not to reuse the name of a scion ....

....collected. The rest of this section analyzes the costs of our protocols and their tolerance to a the fault conditions that are common in actual distributed systems. Memory usage assumes stock 32 bit hardware. 4. 1 Failures The failure assumptions are weaker than in most comparable material [4, 5, 6, 15, 16, 21]. Messages may be duplicated, lost or delivered out of order. Processor pairs are subject to periods during which communication between them may be impossible; however, it is not assumed that this failure is either symmetric or transitive. By minimizing the assumptions we ensure our mechanisms are ....

[Article contains additional citation context not shown here]

Peter William Dickman. Distributed Object Management in a Non-Small Graph of Autonomous Networks with Few Failures. PhD thesis, Darwin College, U. of Cambridge, Cambridge, England (GB), March 1992.

....a cyclic one. Usually, the cyclic distributed garbage collector is triggered at a low rate and most of garbage is assumed to be reclaimed by the acyclic one. However, to be efficient such combination relies on the assumption that global tracing frequency is low compared to the acyclic collector. Dickman [1991] combines his Optimised Weighted Reference Counting with an (unspecified) cyclic global garbage collector. This cyclic collector is responsible for both reclaiming distributed cycles and objects Optimised Weighted Reference Counting can t collect (explained in Section 3.4) Triggering the cyclic ....

....C 1 1 1 t u y t y t 1 b B z A Figure 17. Tracing within a group composed of spaces A and B. main distributed GC characteristics Distributed GC cyclic floating large imple space message failure Techniques garbage scale mented failure duplic. loss late Reference Counting [Bevan, 1987] ffl [Dickman, 1991] ffi ffl ffl ffl [Piquer, 1991] ffl ffl ffl ffl [Goldberg, 1989] ffl Reference Listing [Shapiro et al. 1990] ffl ffl ffl ffl ffl ffl ffl [Birrell et al. 1993] ffl ffl ffl ffi ffl ffl ffl Tracing [Ali, 1984] ffl [Hughes, 1985] ffl ffl [Ladin and Liskov, 1992] ffl ffl [Lang et al. 1992] ....

Peter Dickman. Distributed Object Management in a Non-Small Graph of Autonomous Networks With Few Failures. PhD thesis, University of Cambridge, United Kingdom, September 1991.

....suggests that the dispersion and retraction mechanism should be considered for object oriented load distribution. It also underlines the importance of locality of reference. 4.2 Bellerophon Bellerophon has been developed as part of a Ph.D. project at the Cambridge University Computer Laboratory [Dickman 92a] to investigate scalable object management mechanisms for object support in networks of autonomous computer systems. 4.2.1 System Overview The Bellerophon system appears to be a single widely distributed pool of objects, where only a part is known to each application. Distribution is ....

Peter Dickman: "Distributed Object Management in a Non-Small Graph of Autonomous networks with Few Failures". Ph.D. Thesis, University of Cambridge Computer Laboratory, 1992.

....are much more numerous than remote calls. Intercepting intra process invocations would also involve changing the language run time system or the compiler. 1 The distributed garbage collection algorithm used in DRASTIC is outside the scope of this paper. The interested reader is referred to [Dic92] SE IH IRT ORT Incoming Reference Outgoing Reference X Ray Process pr x IRT : Incoming Reference Table ORT : Outgoing Reference Table IH : InHandler SE : StoreExemplar Fig. 5. A Process Main Data Structures CallHandlers are objects that are provided automatically by the DRASTIC platform, ....

....the only such reference and it is dropped by the object holding it, at the next local GC the ORTSlot s finalize method will be called. This method has been overridden to call DRASTIC s distributed garbage collector, which is based on distributed weighted reference counting and builds on work in [Dic92] When the finalize method is called it triggers DRASTIC platform code which invokes a method on the remote InHandler object, passing the contents of the ORTSlot and requesting that the weight for this object reference be removed. Code at the remote InHandler then removes the necessary weight ....

P. Dickman. Distributed Object Management in a Non-Small Graph of Autonomous Networks With Few Failures. PhD thesis, University of Cambridge, September 1992.

....above, by relying on the existence of a global garbage collector (which is necessary anyway to collect distributed cycles of garbage, since they are not removed by the protocol presented in this paper) to detect and remove garbage scions retained in this way. This is the approach taken by Dickman [7] and means that the algorithm is no longer live, but remains efficient and effective. An alternative would be to apply a rule, similar to the rule for objects, to broken chains: any reference chain indirecting through a terminated space is deemed dangling. Such an approach is only correct, ....

Peter W. Dickman. Distributed Object Management in a Non-Small Graph of Autonomous Networks with Few Failures. PhD thesis, University of Cambridge Computer Laboratory, 1992.

.... and their transmission by a remote procedure call (RPC) layer (e.g. 4] whereas for persistent objects, such an invocation may lead to the object being mapped into an address space from secondary storage (e.g. 2] Such references have been offered in a large number of systems, including [2, 7, 10, 15]. The principal motivation for such references is that they can be used to provide a number of transparency mechanisms [12] In particular, the location of objects, how they are accessed and whether or not they are persistent are all normally concealed from the client. Furthermore, because of ....

....since large amounts of unnecessary error handling code may be produced. It follows that optimisations which ensure that calls are local have benefits in terms of program simplification as well as run time efficiency. In the UTOR model special support, such as the Bellerophon colocator references [7], is required to achieve this. The fragmented object model permits the designer of an FO class to handle such problems within the fragment, however this will not always be possible. If network errors cannot be entirely masked, the FO designer must include the additional exceptions in the interface ....

Peter Dickman. Distributed Object Management in a Non-Small Graph of Autonomous Networks with Few Failures. PhD thesis, University of Cambridge Computer Laboratory, 1992.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC