Distributed Strategic Interleaving with Load Balancing

Transcription

1 Distributed Strategic Intereaving with Load Baancing J.A. Bergstra 1,2 and C.A. Middeburg 1,3 1 Programming Research Group, University of Amsterdam, P.O. Box 41882, 1009 DB Amsterdam, the Netherands 2 Department of Phiosophy, Utrecht University, P.O. Box 80126, 3508 TC Utrecht, the Netherands 3 Computing Science Department, Eindhoven University of Technoogy, P.O. Box 513, 5600 MB Eindhoven, the Netherands janb@science.uva.n,keesm@win.tue.n Abstract. In a previous paper, we deveoped an agebraic theory of threads, intereaving of threads, and interaction of threads with services. In the current paper, we assume that the threads and services are distributed over the nodes of a network. We extend the theory deveoped so far to the distributed case by introducing distributed intereaving strategies that support expicit thread migration and see to oad baancing or capabiity searching by impicit thread migration. Keywords: thread agebra, distributed strategic intereaving, thread migration, oad baancing, capabiity searching CR Categories: C.2.4, D.1.3,.1.2, Introduction The objective of this paper is to deveop an agebraic theory of threads, intereaving of threads, and interaction of threads with services that covers the case where the threads and services are distributed over the nodes of a network and both expicit thread migration and impicit thread migration take pace. A thread is the behaviour of a deterministic sequentia program under execution. Muti-threading refers to the concurrent existence of severa threads in a program under execution. Muti-threading is the dominant form of concurrency provided by contemporary object-oriented programming anguages such as Java [14] and C# [15]. Arbitrary intereaving, on which theories about concurrent processes such as ACP [1, 11], CCS [17] and CSP [16] are based, is not an appropriate abstraction when deaing with muti-threading. In the case of mutithreading, some deterministic intereaving strategy is used. In [7], we introduced a number of pausibe deterministic intereaving strategies for muti-threading. We proposed to use the phrase strategic intereaving for the more constrained form of intereaving obtained by using such a strategy. We aso introduced a feature for interaction of threads with services. The agebraic theory of threads,

2 intereaving of threads, and interaction of threads with services is caed thread agebra. In the current paper, we assume that the threads and services are distributed over the nodes of a network. We extend the theory deveoped so far to the distributed case by introducing distributed intereaving strategies that support expicit thread migration and see to oad baancing or capabiity searching by impicit thread migration. These distributed intereaving strategies are adaptations of the simpest intereaving strategy introduced in [7], to wit pure cycic intereaving, to the distributed case. The other intereaving strategies from [7] can be adapted in a simiar way. Load baancing and capabiity searching are considered impicit forms of thread migration that may be part of distributed intereaving strategies. Load baancing is intended for equaizing the execution speed of the different threads that exist concurrenty in a distributed system. Capabiity searching is intended for achieving that the different threads that exist concurrenty in a distributed system are executed where the services needed are avaiabe. A main assumption made in this paper is that, in the case where systems are networks of nodes over which threads and services are distributed, a singe intereaving strategy determines how the threads that exist concurrenty at the different nodes are intereaved. This is a drastic simpification, as a resut of which intuition may break down. We beieve however that some such simpification is needed to obtain a manageabe theory about the behaviour of such systems and that the resuting theory wi sometimes be adequate and sometimes be inadequate. Moreover, cycic intereaving is a simpification of the intereaving strategies actuay used. Because of the compexity of those strategies, we consider a simpification ike this one desirabe to start with. Our work on thread agebra marks itsef off from much work on the subject of muti-threading because we abandon the usua point of view that arbitrary intereaving is an appropriate abstraction when deaing with muti-threading. The foowing points iustrate why that point of view is untenabe: (i) whether the intereaving of certain threads eads to deadock depends on the intereaving strategy used; (ii) sometimes deadock takes pace with a particuar intereaving strategy whereas arbitrary intereaving woud not ead to deadock and vice versa. We give demonstrations of (i) and (ii) in [7] and [5], respectivey. The thread-service dichotomy made in thread agebra is usefu for the foowing reasons: (i) for services, a state-based description is generay more convenient than an action-based description whereas it is the other way round for threads; (ii) the interaction of threads with services is of an asymmetric nature. In [5], evidence of both (i) and (ii) is produced by the estabished connections of threads and services with processes as considered in an extension of ACP with conditions introduced in [4]. Thread agebra is a design on top of an agebraic theory of the behaviour of deterministic sequentia programs under execution introduced in [3] under the name Basic Poarized Process Agebra (BPPA). Prompted by the deveopment of thread agebra, it has been renamed to Basic Thread Agebra (BTA). 2

3 This paper is organized as foows. Before we take up intereaving strategies for threads that are distributed over the nodes of a network, we review BTA and guarded recursion in the setting of BTA (Sections 2 and 3). We aso discuss the approximation induction principe in this setting (Section 4). After that, we introduce a very simpe intereaving strategy for threads that are distributed over the nodes of a network which supports expicit thread migration (Section 5). Then, we provide a cassification of services that wi be used in subsequent sections (Section 6). oowing this, we introduce a variation of the very simpe intereaving strategy that prevents threads from migrating whie they keep ocks on shared services (Section 7). Thereupon, we introduce three variations of the second intereaving strategy that see to oad baancing by means of impicit migration (Sections 8, 9 and 10). Next, we introduce a variation of the second intereaving strategy that takes care of capabiity searching by means of impicit migration (Section 11). After that, we introduce thread-service composition to aow for threads to be affected by services (Section 12). Then, because it adds to the need for oad baancing, we introduce a basic form of thread forking (Section 13). inay, we make some concuding remarks and mention some options for future work (Section 14). 2 Basic Thread Agebra In this section, we review BTA (Basic Thread Agebra), introduced in [3] under the name BPPA (Basic Poarized Process Agebra). BTA is a form of process agebra which is taiored to the description of the behaviour of deterministic sequentia programs under execution. The behaviours concerned are caed threads. In BTA, it is assumed that there is a fixed but arbitrary set of basic actions A with tau A. We write A τ for A {tau}. The members of A τ are referred to as actions. The intuition is that each basic action performed by a thread is taken as a command to be processed by a service provided by the execution environment of the thread. The processing of a command may invove a change of state of the service concerned. At competion of the processing of the command, the service produces a repy vaue. This repy is either T or and is returned to the thread concerned. Athough BTA is one-sorted, we make this sort expicit. The reason for this is that we wi extend BTA with additiona sorts in Section 5. The agebraic theory BTA has one sort: the sort T of threads. BTA has the foowing constants and operators: the deadock constant D : T; the termination constant S : T; for each a A τ, the binary postconditiona composition operator a : T T T. Terms of sort T are buit as usua. Throughout the paper, we assume that there are infinitey many variabes of sort T, incuding x, y, z. 3

4 Tabe 1. Axiom of BTA x tau y = x tau x T1 Tabe 2. Axioms for guarded recursion X E = t X E if X =t X E RDP E X = X E if X V(E) RSP We use infix notation for postconditiona composition. We introduce action prefixing as an abbreviation: a p, where p is a term of sort T, abbreviates p a p. Let p and q be cosed terms of sort T and a A τ. Then p a q wi perform action a, and after that proceed as p if the processing of a eads to the repy T (caed a positive repy) and proceed as q if the processing of a eads to the repy (caed a negative repy). The action tau pays a specia roe. Its execution wi never change any state and aways ead to a positive repy. BTA has ony one axiom. This axiom is given in Tabe 1. Using the abbreviation introduced above, axiom T1 can be written as foows: x tau y = tau x. The structura operationa semantics of BTA can be found in [7, 8]. 3 Guarded Recursion We proceed to recursion in the current setting. We extend BTA with guarded recursion by adding constants for soutions of guarded recursive specifications and axioms concerning these additiona constants. We wi write BTA+REC for the resuting theory. A recursive specification over BTA is a set of equations E = {X = t X X V }, where V is a set of variabes of sort T and each t X is a term of sort T from the anguage of BTA+REC that ony contains variabes from V. We write V(E) for the set of a variabes that occur on the eft-hand side of an equation in E. Let t be a term of sort T from the anguage of BTA+REC containing a variabe X of sort T. Then an occurrence of X in t is guarded if t has a subterm of the form t a t containing this occurrence of X. A recursive specification E is guarded if a occurrences of variabes in the right-hand sides of its equations are guarded or it can be rewritten to such a recursive specification using the equations of E. We are ony interested in modes of BTA in which guarded recursive specifications have unique soutions. or each guarded recursive specification E and each X V(E), we add a constant of sort T standing for the unique soution of E for X to the constants of BTA. The constant standing for the unique soution of E for X is denoted by X E. Moreover, we add the axioms for guarded recursion given in Tabe 2 to BTA, where we write t X E for t X with, for a Y V(E), a occurrences of Y in t X repaced by Y E. In this tabe, X, t X and E stand for an arbitrary variabe of sort T, an arbitrary term of sort T from the anguage of BTA+REC and an 4

5 Tabe 3. Approximation induction principe n 0 πn(x) = πn(y) x = y AIP Tabe 4. Axioms for projection operators π 0 (x) = D π n+1 (S) = S π n+1 (D) = D P0 P1 P2 π n+1 (x a y) = π n(x) a π n(y) P3 arbitrary guarded recursive specification over BTA, respectivey. Side conditions are added to restrict the variabes, terms and guarded recursive specifications for which X, t X and E stand. The equations X E = t X E for a fixed E express that the constants X E make up a soution of E. The conditiona equations E X = X E express that this soution is the ony one. RDP stands for recursive definition principe and RSP stands for recursive specification principe. 4 Approximation Induction Principe Each cosed term of sort T from the anguage of BTA denotes a finite thread, i.e. a thread of which the ength of the sequences of actions that it can perform is bounded. However, not each cosed term of sort T from the anguage of BTA+REC denotes a finite thread: recursion gives rise to infinite threads. Cosed terms of sort T from the anguage of BTA+REC that denote the same infinite thread cannot aways be proved equa by means of the axioms of BTA+REC. In this section, we introduce the approximation induction principe to remedy this. The approximation induction principe, AIP in short, is based on the view that two threads are identica if their approximations up to any finite depth are identica. The approximation up to depth n of a thread is obtained by cutting it off after performing a sequence of actions of ength n. AIP is the infinitary conditiona equation given in Tabe 3. Here, foowing [3], approximation of depth n is phrased in terms of a unary projection operator π n. The axioms for the projection operators are given in Tabe 4. In this tabe, a stands for an arbitrary member of A τ. Let T stand for either BTA, BTA+REC or any other extension of BTA introduced in this paper. Then we wi write T +AIP for T extended with the projections operators π n and the axioms P0 P3 and AIP. RDP, RSP and AIP originate from work on ACP [1]. In the setting of ACP, these principes were first formuated in [2]. Like in the setting of ACP, RSP foows from RDP and AIP. 5

6 5 Distributed Strategic Intereaving of Threads In this section, we take up the extension of BTA to a theory about threads and distributed muti-threading by introducing a very simpe distributed intereaving strategy. The resuting theory is caed TA dsi. In order to dea with threads and services that are distributed over the nodes of a network, it is assumed that there is a fixed but arbitrary finite set L of ocations such that L N. The set LA of ocated basic actions is defined by LA = {.a L a A}. Henceforth, basic actions wi aso be caed unocated basic actions. We write LT for the set {.tau L}. The members of LA LT are referred to as ocated actions. Performing an unocated action a is taken as performing a at a ocation sti to be fixed by the distributed intereaving strategy. Performing a ocated action.a is taken as performing a at ocation. Performing an action a at ocation brings with it processing of a by a service at ocation. Threads that perform unocated actions ony are caed unocated threads and threads that perform ocated actions ony are caed ocated threads. The behaviour of a distributed system of the kind considered in this paper is supposed to be a ocated thread. However, the threads that exist concurrenty at the different ocations in the system are unocated threads. It is assumed that the coection of a threads that exist concurrenty at the same ocation takes the form of a sequence of unocated threads, caed the oca thread vector at the ocation concerned. It is assumed that the coection of oca thread vectors that exist concurrenty at the different ocations takes the form of a sequence of pairs, one for each ocation, consisting of a ocation and the oca thread vector at that ocation. Such a sequence is caed a distributed thread vector. Distributed strategic intereaving operators turn a distributed thread vector of arbitrary ength into a singe ocated thread. This singe ocated thread obtained via a distributed strategic intereaving operator is aso caed a distributed muti-thread. In this paper, we cover adaptations of the simpest intereaving strategy, namey pure cycic intereaving, which support expicit thread migration and see to oad baancing or capabiity searching by impicit thread migration. In the distributed case, cycic intereaving basicay operates as foows: at each stage of the intereaving, the first thread in the first oca thread vector in the distributed thread vector gets a turn to perform an action at the ocation at which the first oca thread vector is and then the distributed thread vector undergoes cycic permutation. We mean by cycic permutation of a distributed thread vector that the first thread in the first oca thread vector becomes the ast thread in the first oca thread vector, a other threads in the first oca thread vector move one position to the eft, the resuting oca thread vector becomes the ast oca thread vector in the distributed thread vector, and a other oca thread vectors in the distributed thread vector move one position to the eft. If a thread in one of the oca thread vectors deadocks, the whoe does not deadock ti a others have terminated or deadocked. An important 6

7 property of cycic intereaving is that it is fair, i.e. there wi aways come a next turn for a active threads. Other pausibe intereaving strategies for the case where threads and services are not distributed are treated in [7]. They can aso be adapted to the case where threads and services are distributed. When discussing intereaving strategies on distributed thread vectors, we use the term current thread to refer to the first thread in the first oca thread vector in a distributed thread vector and we use the term current ocation to refer to the ocation at which the first oca thread vector in a distributed thread vector is. TA dsi has the sort T of BTA and in addition the foowing sorts: the sort LT of ocated threads; the sort LTV of oca thread vectors; the sort DTV of distributed thread vectors. To buid terms of sort T, TA dsi has the constants and operators of BTA and in addition the foowing operators: for each n N, the binary migration postconditiona composition operator mg(n) : T T T. To buid terms of sort LT, TA dsi has the foowing constants and operators: the deadock constant D : LT; the termination constant S : LT; for each L and a A τ, the binary postconditiona composition operator.a : LT LT LT; the unary deadock at termination operator S D : LT LT; the unary cycic distributed strategic intereaving operator :DTV LT. To buid terms of sort LTV, TA dsi has the foowing constants and operators: the empty oca thread vector constant : LTV; the unary singeton oca thread vector operator : T LTV; the binary oca thread vector concatenation operator :LTV LTV LTV. To buid terms of sort DTV, TA dsi has the foowing constants and operators: the empty distributed thread vector constant : DTV; for each L, the unary singeton distributed thread vector operator [ ] : LTV DTV; the binary distributed thread vector concatenation operator : DTV DTV DTV. Terms of the different sorts are buit as usua for a many-sorted signature (see e.g. [18, 20]). Throughout the paper, we assume that there are infinitey many variabes of sort LT, incuding u, v, w, infinitey many variabes of sort LTV, incuding α, α 1, α 2, and infinitey many variabes of sort DTV, incuding β, β 1, β 2. 7

8 We introduce ocated action prefixing as an abbreviation:.a p, where p is a term of sort LT, abbreviates p.a p. The overoading of D, S, and coud be resoved, but we refrain from doing so because it is aways cear from the context which constant or operator is meant. Essentiay, the sort DTV incudes a sequences of pairs consisting of a ocation and a oca thread vector. The ones that contain a unique pair for each ocation are the proper distributed thread vectors in the sense that the cycic distributed intereaving strategy outined above is intended for them. Improper distributed thread vectors that do not contain dupicate pairs for some ocation are needed in the axiomatization of this strategy. Improper distributed thread vectors that do contain dupicate pairs for some ocation appear to have more than one oca thread vector at the ocation concerned. Their excusion woud make it necessary for concatenation of distributed thread vectors to be turned into a partia operator. The cycic distributed intereaving strategy never turns a proper distributed thread vector into an improper one or the other way round. Simiar remarks appy to the enriched distributed thread vectors introduced in subsequent sections and wi not be repeated. Let p and q be cosed terms of sort LT from the anguage of TA dsi, L and a A τ. Then p.a q wi perform action a at ocation, and after that proceed as p if the processing of a eads to the repy T (caed a positive repy) and proceed as q if the processing of a eads to the repy (caed a negative repy). The deadock at termination operator S D is an auxiiary operator used in the axioms for cycic distributed intereaving introduced beow to turn termination into deadock. Let p be a cosed term of sort LT from the anguage of TA dsi. Then S D (p) behaves the same as p, except that it deadocks whenever p woud terminate. The cycic distributed strategic intereaving operator serves for intereaving of the threads in a proper distributed thread vector according to the strategy outined above, but with support of expicit thread migration. In the case where a oca thread vector of the form p mg(n) q α with n L is encountered as the first oca thread vector, α becomes the ast oca thread vector in the distributed thread vector and p is appended to the oca thread vector at ocation n. If n L, then α q becomes the ast oca thread vector in the distributed thread vector. The migration postconditiona composition operators have the same shape as the postconditiona composition operators introduced earier. However, formay no action is invoved in migration postconditiona composition. Notice that, in case of improper distributed thread vectors, threads that are expicity migrated to some ocation may evaporate or appear in the first of severa oca thread vectors at the ocation concerned. In the axioms for cycic distributed intereaving introduced beow, a binary function app ( L) from unocated threads and distributed thread vectors to distributed thread vectors is used which maps each unocated thread x and distributed thread vector β to the distributed thread vector obtained by appending 8

9 Tabe 5. Definition of the functions app app (x, ) = app (x, [α] β) = [α x ] β if = app (x, [α] β) = [α] app (x, β) if Tabe 6. Axioms for postconditiona composition u t v = u t u LT1 Tabe 7. Axioms for deadock at termination S D (S) = D S2D1 S D (D) = D S2D2 S D (.tau u) =.tau S D (u) S2D3 S D (u.a v) = S D (u).a S D (v) S2D4 Tabe 8. Axioms for cycic distributed intereaving ( ) = S CDI1 ([ ] 1... [ ] k ) = S CDI2 ([ ] β) = (β [ ] ) CDI3 ([ S α] β) = (β [α] ) CDI4 ([ D α] β) = S D ( (β [α] )) CDI5 ([ tau x α] β) =.tau (β [α x ] ) CDI6 ([ x a y α] β) = (β [α x ] ).a (β [α y ] ) CDI7 ([ x mg(n) y α] β) =.tau (app n (x, β [α] )) if n L CDI8 ([ x mg(n) y α] β) =.tau (β [α y ] ) if n L CDI9 x to the oca thread vector at ocation in β. The functions app are defined in Tabe 5. The axioms for postconditiona composition on ocated threads, deadock at termination and cycic distributed intereaving are given in Tabes 6, 7 and 8. In these tabes, t stands for an arbitrary action from LT, 1,..., k and stand for arbitrary ocations from L, a stands for an arbitrary action from A, and n stands for an arbitrary natura number. Axioms CDI1 CDI9 set out the detais of the simpe cycic distributed intereaving strategy introduced in this section in a cear and concise way. To be fuy precise, we shoud give axioms concerning the constants and operators to buid terms of the sorts LTV and DTV as we. We refrain from doing so because the constants and operators concerned are the usua ones for sequences the singeton distributed thread vector operators invove an impicit pairing of their operand with a ocation. 4 Simiar remarks appy to the sorts 4 This amounts to ooking upon [α] as syntactic sugar for (, α). 9

10 of oca thread vectors and distributed thread vectors introduced in subsequent sections and wi not be repeated. Guarded recursion can be added to TA dsi as it is added to BTA in Section 3. We wi write TA dsi +REC for the resuting theory. The axioms for cycic distributed strategic intereaving given in Tabe 8 constitute a definition by recursion on the sum of the depths of a threads in the distributed thread vector with case distinction on the structure of the first thread in the first oca thread vector. Hence, it is obvious that the axioms are consistent and that every cosed term of the sort LT from the anguage of TA dsi is derivaby equa to one that does not contain the cycic distributed strategic intereaving operator. Moreover, it foows easiy that every projection of a cosed term of the sort LT from the anguage of TA dsi +REC is derivaby equa to one that does not contain the cycic distributed strategic intereaving operator. Simiar remarks appy to the other distributed strategic intereaving operators introduced in this paper and wi not be repeated. We refrain from giving the structura operationa semantics of TA dsi. The transition rues for the postconditiona composition operators on ocated threads constitute a trivia adaptation of the transition rues for the postconditiona composition operators on unocated threads. The transition rues for the deadock at termination operator can be found in [7, 8]. The transition rues for the cycic distributed strategic intereaving operator, as we as the transition rues for the distributed strategic intereaving operators that wi be introduced ater in this paper, are simiar to the transition rues for the strategic intereaving operator for cycic intereaving found in [7, 8]. However, the transition rues for a distributed strategic intereaving operator are so much more invoved that they might be caed unreadabe. This impies that they cannot add to a better understanding of the distributed intereaving strategy concerned. Henceforth, we wi write TA dsi (A) for TA dsi with the set of basic actions A fixed to be the set A. 6 A Cassification of Services In this section, we provide a usefu cassification of services present at each of the ocations over which the threads and services of a system are distributed. This cassification is a sighty adapted version of the cassification given in [7] and wi be used in subsequent sections. A major distinction is between target services and para-target services: A service is a target service if the resut of the processing of commands by the service is party observabe externay. Printing a document, sending an emai message, showing data on a dispay or writing persistent data in permanent memory are typica exampes of using a target service. A service is a para-target service if the resut of the processing of commands by the service is whoy unobservabe externay. Setting a timer or transferring data by means of a Java pipe are typica exampes of using a para-target service. 10

11 In the case of para-target services, a further distinction is between private services and shared services: A private service provides its service to a singe thread ony. A timer is usuay a private service. Private services exist to support a singe thread in creating intended behaviour in reation to target services. A shared service provides its service to a threads existing at some ocation. A Java pipe is an exampe of a shared para-target service. Shared services exist to support a threads existing at the same ocation in creating intended behaviour in reation to target services. To simpify matters, it wi be assumed that a para-target services at some ocation are either private or shared, thus disregarding the possibiity that a service provides its service to some of the threads existing at the ocation ony. This eaves us with execution architectures that provide target services, shared services and private services. Private services are caed oca services in [7]. We use the phrase private service here because oca service may be confusing in the current setting where services are ocated. The overa intuition about threads, target services and para-target services is that: a thread is the behaviour of a sequentia deterministic program; a distributed muti-threaded system consists of a number of threads, which are intereaved according to some deterministic distributed intereaving strategy and which interact with a number of services; the intentions about the resuting behaviour pertain ony to interaction with target services; interaction with para-target services takes pace ony in as far as it is needed to obtain the intended behaviour in reation to target services. One of the assumptions made in thread agebra is that para-target services are deterministic. The excusion of non-deterministic para-target services, ike the excusion of non-deterministic intereaving strategies, is a simpification. We beieve however that this simpification is adequate in the cases that we address: services that keep private data for a thread or shared data for a number of threads. Of course, it is inadequate in cases where services such as dice-paying services are taken into consideration. Another assumption is that target services are non-deterministic. The reason for this assumption is that the dependence of target services on externa conditions make it appear to threads that they behave non-deterministicay. Henceforth, to simpify matters, it is assumed that each thread has a unique private service. This service has to move aong with the thread if it migrates to another ocation. 11

12 7 Distributed Strategic Intereaving and Locking In this section, we introduce a variation of the distributed intereaving strategy from Section 5 that prevents threads from migrating whie they keep ocks on shared services. This resuts in a theory caed TA ck dsi. In order to dea with ocking of shared services, it is assumed that there is a fixed but arbitrary finite set of foci and a fixed but arbitrary finite set of methods M. Each focus pays the roe of a name of a service provided by the execution environment that can be requested to process a command. Each method pays the roe of a command proper. We write M for the set {f.m f, m M}. or the set A of unocated actions, we take the set M. Performing an unocated action f.m is taken as making a request to the service named f at a ocation sti to be fixed by the distributed intereaving strategy to process command m. Performing a ocated action.f.m is taken as making a request to the service named f at ocation to process command m. It is aso assumed that t, s 1, s 2,.... The foci t, s 1, s 2,... pay a specia roe: for each thread, t is the focus of the unique private service of that thread; for each ocation, s 1, s 2,... are the foci of the services shared by a threads from the oca thread vector at that ocation. We write s for the set {s 1, s 2,...} of shared service foci. or each ocation, a threads from the oca thread vector at that ocation are supposed to be abe to ock and unock the services with foci from s at that ocation. It is assumed that ock, unock M. We write M for the set {ock, unock} of ocking methods. When processing a command, the services with foci from s do nothing but changing a state and producing repies on the basis of the state. In every state, either the processing of ock eads to a positive repy or the processing of unock eads to a positive repy, but not both, and if the processing of one of them eads to a positive repy, then the processing concerned moves the service to a state in which the processing of the other one eads to a positive repy. In this way, successive ockings without intermediate unocking is prevented. It is assumed that initiay the processing of ock eads to a positive repy. Moreover, it is assumed that a request to process ock or unock is rejected by services with foci other than foci from s. Rejecting a request to process a command brings with it deadocking of the thread making the request. If a thread successfuy performs the action s i.ock, it acquires the ock of the service with focus s i. It keeps the ock unti it is reeased by performing s i.unock. A threads are supposed to work as foows. or each focus s i, a commands s i.m different from s i.ock must be performed in a phase in which the thread keeps the ock of the service with focus s i. If a threads adhere to this rue, it is guaranteed that a thread keeping the ock on a service has excusive access to that service. The distributed intereaving strategy introduced in this section permits of no deviation from this rue. 12

13 Tabe 9. Definition of the functions app app (x, ) = app (x, [α] β) = [α x ] β if = app (x, [α] β) = [α] app (x, β) if In the presence of ocking and unocking of shared services, axiom CDI8 (Tabe 8) is not satisfactory. A thread from the oca thread vector at one ocation shoud not be permitted to migrate to another ocation if the thread keeps the ock of one or more shared services at the former ocation. To dea with that, we have to enrich distributed thread vectors by repacing each thread in each oca thread vector by a pair consisting of the thread and the set of a foci naming shared services on which the thread keeps a ock. We mention that, when the intereaving of the threads in a distributed thread vector makes a start, the threads are supposed to keep no ock. TA ck dsi has the same sorts as TA dsi(m). To buid terms of the sorts T, LT and DTV, TA ck dsi has the same constants and operators as TA dsi(m). To buid terms of sort LTV, TA ck dsi has the foowing constants and operators: the empty oca thread vector constant : LTV; for each s, the unary singeton oca thread vector operator : T LTV; the binary oca thread vector concatenation operator :LTV LTV LTV. That is, the operator is repaced by the operators. Simiar to the singeton distributed thread vector operators [ ], the singeton oca thread vector operators invove an impicit pairing of their operand with a set of foci naming shared services. In the axioms for cycic distributed intereaving with ocking introduced beow, a binary function app ( L) from unocated threads and distributed thread vectors to distributed thread vectors is used which maps each unocated thread x and distributed thread vector β to the distributed thread vector obtained by appending x to the oca thread vector at ocation in β and associating the empty set as set of foci with x. The functions app are defined in Tabe 9. The axioms for cycic distributed intereaving with ocking are given in Tabe 10. In this tabe, 1,..., k and stand for arbitrary ocations from L, f stands for an arbitrary focus from, m stands for an arbitrary method from M, stands for an arbitrary subset of s, and n stands for an arbitrary natura number. Axiom CDI7 of the strategy from the previous section is repaced by four axioms to make distinction between ock actions, unock actions and other actions. This distinction must be made to keep track of the ocks that the different threads keep and to precude the possibiity of deviation from the rue that threads must adhere to in order to guarantee excusive access to ocked services. Guarded recursion can be added to TA ck dsi as it is added to BTA in Section 3. 13

14 Tabe 10. Axioms for cycic distributed intereaving with ocking ( ) = S CDIck1 ([ ] 1... [ ] k ) = S CDIck2 ([ ] β) = (β [ ] ) CDIck3 ([ S α] β) = (β [α] ) CDIck4 ([ D α] β) = S D ( (β [α] )) CDIck5 ([ tau x α] β) =.tau (β [α x ] ) CDIck6 ([ x f.m y α] β) = (β [α x ] ).f.m (β [α y ] ) if f s (f m M ) CDIck7a ([ x f.ock y α] β) = (β [α x {f} ] ).f.ock (β [α x f.ock y ] ) if f s \ CDIck7b ([ x f.unock y α] β) = (β [α x \{f} ] ).f.unock (β [α y ] ) if f CDIck7c ([ x f.m y α] β) = S D ( (β [α] )) if f s (f m ock) (f s \ m = ock) CDIck7d ([ x mg(n) y α] β) =.tau (app n (x, β [α] )) if n L = CDIck8 ([ x mg(n) y α] β) =.tau (β [α y ] ) if n L CDIck9 8 Distributed Strategic Intereaving and Load Baancing In this section, we introduce a variation of the distributed intereaving strategy from Section 7 with impicit migration of threads to achieve oad-baancing. This resuts in a theory caed TA ba dsi. Immediatey after the current thread has performed an action, impicit migration of that thread to another ocation may take pace. Whether migration reay takes pace, depends on three parameters of the strategy: k 1, k 2 and k 3. A three parameters are natura numbers greater than 0. The current thread is impicity migrated if the foowing conditions are fufied: the current thread is migration compatibe; the current thread keeps no ocks; the number of steps that the current thread has performed at its current ocation is at east k 1 ; the argest number of steps that the current thread wi perform before termination, deadock or expicit migration is at east k 2 ; the ength of the oca thread vector that contains the current thread, say n, is greater than k 3 and there is another oca thread vector whose ength times 2 is ess than n. The purpose of these conditions is as foows: 14

15 the first condition is to excude migration of threads whose interaction with services is in whoe or in part bound to services at the ocations they are; the second condition is to prevent threads from migrating whie they keep ocks on shared services; the third condition is to guard threads against hopping around useessy; the fourth condition is to precude the possibiity that threads are migrated whie they wi necessariy terminate, deadock or migrate expicity soon; the ast condition is to achieve that threads from ocations with ong oca thread vectors are migrated to ocations with short oca thread vectors. If the current thread is impicity migrated, then it wi be migrated to the first among the ocations with a shortest oca thread vector. To understand why oad baancing strategic intereaving may be usefu, assume that: (i) time is divided into sices of equa ength; (ii) during each sice, a oca thread vectors get a turn to perform an action of one thread. Thus, a ong oca thread vector gives rise to sow execution per thread whereas a short oca thread vector gives rise to fast execution per thread. Load baancing wi equaize the execution speed of the threads in the different oca thread vectors. To dea with both ocking and oad baancing, we have to enrich distributed thread vectors by repacing each thread in each oca thread vector by a tripe consisting of the thread, the set of a foci naming services on which the thread keeps a ock, and the minimum of k 1 and the number of steps that the thread has performed at the ocation it is. We mention that, when the intereaving of the threads in a distributed thread vector makes a start, the threads are supposed to keep no ock and to have performed no steps at the ocation they are. It is assumed that each method can be cassified as either functiona or nonfunctiona, so that the set M fn M of functiona methods can be distinguished. Moreover, it is assumed that ock M fn and unock M fn. A method is a functiona method if its processing by a shared service never change the state of the shared service. Migration compatibiity is a criterion to excude migration of threads whose interaction with services is in whoe or in part bound to services at the ocation they are. Interaction of threads with target services is aways bound, interaction of threads with their private service is never bound, and interaction of threads with shared services using functiona methods ony is never bound. However, interaction of threads with shared services may be bound if non-functiona methods are used. In this paper, a thread is considered migration compatibe if its interaction with services incudes nothing but interaction with private services and interaction with shared services using functiona methods. This makes migration compatibiity rather restrictive. We can take a ess restrictive criterion instead if we assume that ocked sessions with shared services can aso be cassified as either functiona or nonfunctiona. The criterion in question is compicated in such a way that it woud distract the attention from the main issues treated in this paper. 15

16 Tabe 11. Definition of the functions app app (x, ) = app (x, [α] β) = [α x 0 ] β if = app (x, [α] β) = [α] app (x, β) if TA ba dsi has the same sorts as TA ck dsi. To buid terms of the sorts T, LT and DTV, TA ba dsi has the same constants and operators as TA ck dsi. To buid terms of sort LTV, TA ba dsi has the foowing constants and operators: the empty oca thread vector constant : LTV; for each s and c N such that c k 1, the unary singeton oca thread vector operator c : T LTV; the binary oca thread vector concatenation operator :LTV LTV LTV. That is, the operator is repaced by the operators c. Simiar to the singeton distributed thread vector operators [ ], the singeton oca thread vector operators c invove an impicit tuping of their operand with a set of foci naming shared services and a natura number ess than or equa to k 1. In the axioms for cycic distributed intereaving with oad baancing introduced beow, a binary function app ( L) from unocated threads and distributed thread vectors to distributed thread vectors is used which maps each unocated thread x and distributed thread vector β to the distributed thread vector obtained by appending x to the oca thread vector at ocation in β and associating the empty set as set of foci and zero as step count with x. The functions app are defined in Tabe 11. Moreover, a unary function pv on distributed thread vectors is used which permutes distributed thread vectors cycicy with impicit migration as outined above. The unary function pv on distributed thread vectors is defined using a number of auxiiary functions: a unary function mgc from unocated threads to Booeans, mapping each unocated thread that is migration compatibe to T and each unocated thread that is not migration compatibe to ; for each n N, a unary function nps n from unocated threads to Booeans, mapping each unocated thread to T if the argest number of steps that it wi perform before termination, deadock or expicit migration is at east n, and to otherwise (nps stands for number of potentia steps); a unary function stv from distributed thread vectors to pairs consisting of a ocation and a oca thread vector, mapping each distributed thread vector to the pair consisting of the first among the ocations with a shortest oca thread vector and the oca thread vector at that ocation; a unary function oc from pairs consisting of a ocation and a oca thread vector to ocations, mapping each such pair to its first component; 16

17 Tabe 12. Definition of the functions mgc n and mgc mgc 0 (x) = T mgc n+1 (S) = T mgc n+1 (D) = T mgc n+1 (tau x) = mgc n (x) mgc n+1 (x f.m y) = mgc n (x) mgc n (y) if f = t (f s m M fn ) mgc n+1 (x f.m y) = if f t (f s m M fn ) mgc n+1 (x mg(n ) y) = T n 0 mgc n (x) = T mgc(x) = T n 0 mgc n (x) = mgc(x) = a unary function tv from pairs consisting of a ocation and a oca thread vector to oca thread vectors, mapping each such pair to its second component; a unary function imc from distributed thread vectors to Booeans, mapping each distributed thread vector that fufis the conditions for impicit migration to T and each distributed thread vector that does not fufi these conditions to. The function mgc in turn is defined using, for each n N, an auxiiary unary function mgc n from unocated threads to Booeans, mapping each unocated thread whose approximation up to depth n is migration compatibe to T and each thread whose approximation up to depth n is not migration compatibe to. The function mgc, as we as the auxiiary functions mgc n, are defined in Tabe 12. The function pv, as we as the auxiiary functions nps n, stv, oc, tv and imc, are defined in Tabe 13. In the definition of stv, 0 stands for a fixed but arbitrary ocation from L. In the definitions of nps n, stv, imc and pv, as we as in many subsequent definitions, we use the usua basic functions on Booeans and natura numbers. Moreover, we write b e, where b is a Booean expression and e is an equation, for the conditiona equation b = T e. A simpe unary function ic on natura numbers is used as we in the axioms for cycic distributed intereaving with oad baancing introduced beow. It is defined by ic(c) = min(c + 1, k 1 ) for a c N. The axioms for cycic distributed intereaving with oad baancing are given in Tabe 14. In this tabe, 1,..., k and stand for arbitrary ocations from L, f stands for an arbitrary focus from, m stands for an arbitrary method from M, stands for an arbitrary subset of s, c stands for an arbitrary natura number ess than or equa to k 1, and n stands for an arbitrary natura number. The differences with the axioms of the strategy from the previous section are mainy found in axioms CDIba6 and CDIba7a CDIba7d. In those axioms, pv is used to achieve the variant of cycic permutation with impicit migration outined above. Guarded recursion can be added to TA ba dsi as it is added to BTA in Section 3. 17

18 Tabe 13. Definition of the functions nps n, stv, oc, tv, imc, and pv nps 0 (x) = T nps n+1 (S) = nps n+1 (D) = nps n+1 (tau x) = nps n (x) nps n+1 (x f.m y) = nps n (x) nps n (y) nps n+1 (x mg(n ) y) = stv( ) = ( 0, ) stv([α] ) = (, α) en(α 1 ) en(α 2 ) stv([α 1 ] 1 [α 2 ] 2 β) = stv([α 1 ] 1 β) en(α 1 ) > en(α 2 ) stv([α 1 ] 1 [α 2 ] 2 β) = stv([α 2 ] 2 β) oc((, α)) = tv((, α)) = α imc( ) = imc([ ] β) = imc([ x k 1 α] β) = mgc(x) nps k 2 (x) en(α) > k 3 2 en(tv(stv(β))) < en(α) imc([ x c α] β) = if c k 1 pv( ) = pv([ ] β) = β [ ] imc([ x c α] β) pv([ x c α] β) = app oc(stv(β)) (x, β [α] ) imc([ x c α] β) pv([ x c α] β) = β [α x c ] The distributed intereaving strategy with impicit migration of which the axioms are given in Tabe 14 does not count faied attempts of a thread to acquire a ock on a shared service as steps performed by the thread. Counting the faied attempts is pausibe as we. However, it resuts in ooking at past steps and future steps from a different ange because the number of potentia faied attempts in the future is generay indefinite. Because it is assumed that ock is a non-functiona method, it is impossibe that a thread is impicity migrated to another ocation after repeated faied attempts of the thread to acquire a ock on a shared service. At first sight, continuing the attemps at another ocation after repeated faied attempts appears to make sense, but by doing so the thread runs the risk of hopping around useessy whie attempting to acquire a ock. Therefore, we have not considered the rather counter-intuitive assumption that ock is a functiona method. Migration compatibiity is defined such that interaction with services after expicit migration is not taken into account, for that interaction is intended to be with services at another ocation. Take a thread that cannot be migrated impicity because of interaction with a target service before termination, deadock or expicit migration. If that thread woud migrate back expicity before the interaction with a target service instead, then it coud be migrated impicity. 18

19 Tabe 14. Axioms for cycic distributed intereaving with oad baancing ( ) = S ([ ] 1... [ ] k ) = S ([ ] β) = (β [ ] ) ([ S c α] β) = (β [α] ) ([ D c α] β) = S D( (β [α] )) ([ tau x c α] β) =.tau (pv([ x ic(c) α] β)) ([ x f.m y c α] β) = (pv([ x ic(c) α] β)).f.m (pv([ y ic(c) α] β)) if f s (f m M ) ([ x f.ock y c α] β) = (β [α x ic(c) {f} ] ).f.ock (pv([ x f.ock y c α] β)) if f s \ ([ x f.unock y c α] β) = (pv([ x ic(c) \{f} α] β)).f.unock (pv([ y ic(c) α] β)) if f ([ x f.m y c α] β) = S D( (β [α] )) if f s (f m ock) (f s \ m = ock) CDIba1 CDIba2 CDIba3 CDIba4 CDIba5 CDIba6 CDIba7a CDIba7b CDIba7c CDIba7d ([ x mg(n) y c α] β) =.tau (app n (x, β [α] )) if n L = CDIba8 ([ x mg(n) y c α] β) =.tau (β [α y c ] ) if n L CDIba9 Thus, oad baancing is feasibe and usefu because expicit migration is present as an option. The foowing might be considered: a thread transformation that inserts oad baancing migration actions to obtain oad baancing triggered by actions instead of the distributed intereaving strategy. However, there appear to be no obvious candidate for such a transformation. It can be done, but the state space of the thread may expode. 9 Load Baancing with Impicit Migration Back In the distributed intereaving strategy with oad baancing from Section 8, the current thread is impicity migrated ony if it is migration compatibe. Athough it is a simpe criterion, migration compatibiity is undecidabe. This makes the distributed intereaving strategy with oad baancing from Section 8 unconvincing. In this section, we introduce a variation of that strategy which does without migration compatibiity. This resuts in a theory caed TA bb dsi. The approach foowed is to precude the need for migration compatibiity on impicit migration of a thread by migrating the thread impicity back before interaction with a target service or interaction with a shared service using a nonfunctiona method can take pace. Because ocking is an instance of interaction with a shared service using a non-functiona method, impicit migration back does not happen whie the thread keeps one or more ocks. 19

20 Immediatey after the current thread has performed an action, impicit migration back may take pace. The current thread is impicity migrated back if the foowing conditions are fufied: the most recent migration of the current thread was an impicit migration; the next action that the current thread wi perform concerns an interaction with a target service or an interaction with a shared service using a nonfunctiona method; the ocation at which the current thread is differs from the ocation from which its odest impicit migration that was not foowed by an expicit migration has taken pace. If the current thread is impicity migrated back, it wi be migrated to the ocation from which its odest impicit migration that was not foowed by an expicit migration has taken pace. Whether impicit migration of the current thread takes pace, depends as before on the three parameters k 1, k 2 and k 3. The current thread is impicity migrated if the foowing conditions are fufied: the second condition for impicit migration back is not fufied; the current thread keeps no ocks; the number of steps that the current thread has performed at its current ocation is at east k 1 ; the argest number of steps that the current thread wi perform before termination, deadock, expicit migration, interaction with a target service or interaction with a shared service using a non-functiona method is at east k 2 ; the ength of the oca thread vector that contains the current thread, say n, is greater than k 3 and there is another oca thread vector whose ength times 2 is ess than n. To dea with impicit migration back, we have to enrich distributed thread vectors by repacing each thread in each oca thread vector by a quadrupe consisting of the thread, the set of a foci naming services on which the thread keeps a ock, the minimum of k 1 and the number of steps that the thread has performed at the ocation it is, and the ocation from which the odest impicit migration of the thread that was not foowed by an expicit migration has taken pace if it exists and the ocation at which the thread is otherwise. We wi use the term migration-back ocation to refer to this ocation. We mention that, when the intereaving of the threads in a distributed thread vector makes a start, the threads are supposed to keep no ock, to have performed no steps at the ocation they are, and to have not been impicity migrated. TA bb dsi has the same sorts as TA ba dsi. To buid terms of the sorts T, LT and DTV, TA bb dsi has the same constants and operators as TA ba dsi. To buid terms of sort LTV, TA bb dsi has the foowing constants and operators: the empty oca thread vector constant : LTV; for each s, c N such that c k 1 and L, the unary singeton oca thread vector operator c, : T LTV; 20