mmnet Summer School Tuesday 20 – Wednesday 21July, 2004, Canterbury Speakers: • • • • • • • 20 February 2004 David Bacon, IBM TJ Watson. Emery Berger, UMass. Robert Berry, IBM Hursley. Hans Boehm, HP. Dave Detlefs, Sun Microsystems. Rick Hudson, Intel. Eliot Moss, UMass. UKC, February 2004 1 Birrell’s Reference Listing Revisited Richard Jones University of Kent R.E.Jones@ukc.ac.uk Luc Moreau University of Southampton Peter Dickman University of Glasgow Outline 1. 2. 3. 4. Distributed reference counting – benefits & issues. Birrell’s algorithm – example. Weaknesses of Birrell’s description. Our approach • • Graphical notation Formalisation & Proof 5. Extensions • • Fault tolerance Optimisation 6. Conclusion. 20 February 2004 UKC, February 2004 3 Problems of a distributed world Concurrency everywhere • must avoid race conditions, etc Communication is costly • changing the reference count of a remote object may cost 10,000 times as much as changing the count of a local object Not easy to get complete knowledge of object graph • synchronisation is expensive Faults everywhere • communications, processes 20 February 2004 UKC, February 2004 4 Terminology Processes: partition computational and storage resources. Messages pass in point-to-point channels between processes. Channels have properties, such as FIFO or lossy. A reference is local if it refers to an object allocated in the same process; alternatively, it is remote (or global). The owner of a reference is the process that initially allocated the object to which the reference refers. 20 February 2004 UKC, February 2004 5 Distributed Reference Counting/Listing Most widely used DGC technique • Maintain a count of remote references to each global object • Reference listing alternative Benefits • Scalable solution • Easy to implement But… • Cannot reclaim garbage cycles • Easy to implement wrong! 20 February 2004 UKC, February 2004 6 Birrell’s algorithm Birrell, Evers, Nelson, Owicki, and Wobber. Distributed Garbage Collection for Network Objects. DEC SRC technical report 116, 1993. Widely used: Modula-3 Network Objects; Java RMI. Based on reference listing, avoids race conditions of naïve implementations, fault tolerance. 20 February 2004 UKC, February 2004 9 Birrell’s description Process P: owner of O Object table Process Q: a client of O Object table w (o) w (o) weak ref concrete O surrogate for O o.dirtySet = {Q,…} Dirty WireRep: Object set: table: identifiers unique maps IDaof ofwirerep processes owner, w(o) plusthat to index the have local of surrogates. object instance at theof the Concrete and surrogate objects. owner. object. •Dirty-set = the o can be removed the object table. •Client invokes surrogate, whosefrom methods perform RPC to owner. •Marshalling •Client has surrogate for o concrete o in object table 20 February 2004 UKC, February 2004 10 P marshalls o to Q P pushes o onto its stack; sends w(o) to Q. 1. Q looks it up in its object table. P O Q Present: use the object w(o)=NIL: surrogate being created; suspend. 2. Absent: enter w(o)=NIL in object table; send dirty(o) to owner(o); 3. Owner adds Q to its dirtySet(o) and dirty(o) returns. 4. Q creates surrogate(o) and adds it to its object table. 5. Q deletes surrogate(o) and sends clean(o) to owner(o). 20 February 2004 UKC, February 2004 11 Dirty calls A B copy ack dirty {A,B} {A} 20 February 2004 UKC, February 2004 Log keep ref on stack copy dirty ack remove from stack 12 Weaknesses Tightly bound to RPC • Acknowledgement mechanism. Implementation specific. • Assumes method invocation pushes arguments onto stack; • Unique surrogate per process (object-listing) Under-specified • Critical sections • Race conditions • Other scenarios Informal proof • Depends on hard-to-formalise aspects (e.g. stack) 20 February 2004 UKC, February 2004 14 Our contribution Novel graphical notation. Formalisation. Discovered requirement for pivotal new states. Proof. 20 February 2004 UKC, February 2004 15 New graphical notation Intuitive. Precise. Uniformity of ‘direction’ of transitions. ‘Obvious’ where transitions are needed. 20 February 2004 UKC, February 2004 16 Lifecycle of references GC unreachable send copy OK RRAR rcv Ack ccitnil ccit Receive reference and note the source ... OK 1 2 3 dirty_ack from Owner send copy_ack to Sender nil Receive reference and note the source nil Obvious where transitions are needed •E.g. Receive reference at state ccit. •ccitnil critical for correctness. 20 February 2004 UKC, February 2004 17 Slicing 20 February 2004 UKC, February 2004 18 Fault tolerance ccitu Slicing OK • Owner is aware we have a reference. ccit ccitl nilu OK ccitnil nil nil ccitnilu nill • Owner is not aware we have a reference. ccitnill 20 February 2004 UKC, February 2004 19 Benefits Intuitive – fault-tolerant version literally encapsulates failure-free version. Identify precisely when failures can be detected. Define states reached after failures detected. Remedial actions. 20 February 2004 UKC, February 2004 20 Formalisation Abstract machine • Processes communicating by asynchronous message passing. • Atomic transitions involve 1 process at a time. Receipt of message changes only a process’ internal state • Trigger sending of a another message? • Store some info in a to do table? 20 February 2004 UKC, February 2004 22 Benefits Inputs and outputs desynchronised. Size of critical sections explicit and minimised. Asynchronous outputs (e.g. background daemon processes to do tables). Suitable for mechanical proof. 20 February 2004 UKC, February 2004 23 Formalisation Rule name: guard pseudo-statements. name guard make_copy (p1,p2,r): p1 p2 receive_T(p1,r)=OK locallyReachable(p1,r) table { id := new Identifier; dirty_T(p1,r) := dirty_T(p1,r) U (p1,p2,id); post(p1, p2, copy(r,id)); } message 20 February 2004 UKC, February 2004 24 More formally Tables defined as functions whose first argument is a process. Channels are bags of messages between pairs of processes. A configuration of the abstract machine is a tuple of all tables and message channels. Pseudo-statements act as configuration transformers: • Given a configuration <…,table_T,…, k>, • table_T(a0,…an):=V denotes <…,table_T',…,k> where table_T'(x0,…xn) = table_T(a0,…an) if (x0,…xn) (a0,…an) table_T'(a0,…an) = V • post(p1,p2,m) denotes <…,table_T,…,k'> where k'(p1,p2) = k(p1,p2) {m} k'(pi,pj) = k(pi,pj), (pi,pj) (p1,p2) 20 February 2004 UKC, February 2004 25 Proof style Safety & Liveness Invariance-based proof • Induction on length of transitions. • Case analysis of transitions. • Termination measure. Benefits • Systematic. • Less error prone than temporal reasoning. – E.g. establishing fine details such as mutual exclusivity complicated in a formalism based on temporal reasoning. 20 February 2004 UKC, February 2004 26 Example proof Lemma: For any processes p1, p2, for any reference r, for any identifier id and for any configuration, the following implication holds: If <p1,p2,id> dirty_T(p1,r) then receive_T (p1,r) = OK Proof: In the initial configuration, dirty tables are empty and the implication trivially holds. We consider the four rules that add/remove entries to/from dirty tables and that modify the content of receive tables to/from OK. • make_copy (p1, p2,r): make_copy adds an entry <p1,p2,id>, and its guard ensures that the receive-table is in the OK state. •… 20 February 2004 UKC, February 2004 27 Key Lemmas Safety Lemma 1: Usable Reference For any processes p1 and p2, for any reference r with p1=owner(r) and p1p2, and for any configuration, the following implication holds: If receive_T(p1,r)=OK, then p1 dirty_T(p2,r). Safety Lemma 2: Reference in Transit For any processes p1, p2, for any reference r, for any identifier id and for any configuration, the following implication holds: permanent If copy(r,id) k(p1,p2), then p1 dirty_T(owner(r),r), if p1 owner(r) temporary or <owner(r),p2,id> dirty_T(owner(r),r), if p1 = owner(t) Safety Lemma 3: Unusable Reference For any process p1, for any reference r and for any configuration, the following implication holds: If receive_T(p1, r)=nil receive_T(p1, r)=ccitnil, then there exists p such that p dirty_T(owner(r), r) or there exist p,id such that <owner(r),p,id> dirty_T(owner(r), r). 20 February 2004 UKC, February 2004 29 Birrell’s algorithm is Safe A DGC algorithm is safe if the collector cannot reclaim live objects. For Birrell's algorithm, there must be an entry in the owner's dirty table for every live object. Birrell's Safety Requirement For all references r, and for all processes p1 and p2 and all identifiers id, If receive t(p1,r)=OK receive_T(p1,r)=nil receive_T(p1,r)=ccitnil copy (r,id) k(p1,p2), then there exists p such that p dirty_T(owner(r),r) or there exist p,id such that <owner(r),p,id> dirty_T(owner(r),r). The proof follows directly from the 3 safety lemmas. 20 February 2004 UKC, February 2004 30 Liveness Liveness guarantees that if all references to an object are deleted, the owner’s dirty table will eventually become empty. To prove this, • We show that whenever there’s a message in a channel, a transition can be fired to consume it. • We introduce a termination measure on the configurations that shows how far the abstract machine is from completing, and show that DGC transitions cause this measure to decrease. • Hence all transition paths terminate. 20 February 2004 UKC, February 2004 31 Termination measures termination_measure(c) = tab_measure + msg_measure(m) + rt_measure(receive_T(p,r)) size of tables messages between pairs of processes tab_measure = 9|dirty_call_todo_T| + 7|dirty_ack_todo_T| + 2|copy_ack_todo_T| + 2|clean_ack_todo_T| + 2|blocked_T| and rt_measure(OK) =5 msg_measure(copy) = 14 rt_measure(ccitnil) =2 msg_measure(dirty) =8 rt_measure(ccit) =1 msg_measure(dirtyack) =6 rt_measure(nil) =1 msg_measure(clean) =3 rt_measure() =0 msg_measure(copyack) =1 msg_measure(cleanack)= 1 states of references in processes values chosen ‘arbitrarily’ 20 February 2004 UKC, February 2004 32 Example: receive_dirty_ack Dmeasure = -1 receive_dirty_ack (p1, p2,r) : dirtyack(r) k (p1, p2) { receive(p1,p2,dirtyack(r)); //-6 copyack_todo_T(p2) := copyack_todo_T(p2) blocked_T(p2,r); //-X // Deserialisation code to be resumed for each entry in blocked_T(p2,r) blocked_T(p2,r) := ; receive_T(p2,r) := OK; //-X //+5 } Thus, termination measure decreases by 1. 20 February 2004 UKC, February 2004 33 Optimisations FIFO channels • Less synchronisation needed • Fewer messages: no clean_ack • Fewer tables. Sender is owner • No need for dirty_call and copy_ack • But need message ordering to avoid races Receiver is owner • Fewer dirty table entries • Again need message ordering 20 February 2004 UKC, February 2004 34 Future work Convince ourselves of appropriateness of Birrell’s remedial actions. Correctness proof of fault-free version. Explore applicability of our techniques • Graphical notation • Proof-techniques • Generality Auto-generation of code from formalism. 20 February 2004 UKC, February 2004 35 Conclusion Intuitive graphical notation. Formal, implementation-independent specification and proof of a widely used algorithm. Discovered weaknesses in original presentation. A widely applicable technique? 20 February 2004 UKC, February 2004 36 Questions? FINIS 20 February 2004 UKC, February 2004 38