Garbage Collecting the World. --Bernard Lang, Christian and Jose

1. Garbage Collecting the World. --Bernard Lang, Christian and Jose Presented by Shikha Khanna coen 317 Date – May25’ 2005

2. Index • Introduction • Terminology • Basic Algorithm • Handling Failures • Group Contention • Modified marking scheme • Conclusion

3. Introduction • Computations performed by collection of processes are more and more common today. Ex - distributed symbolic computations - distributed databases This involves existence of remote references i.e objects at distant node referencing memory in each others address space which leads to unused memory. • The paper presents a distributed garbage collection algorithm to remove such unused memory.

4. Terminology • Node – Processor or a process on a processor able to manage its own memory space. • Mutator – process that allocates chunks of memory (cells). Cells contain references to cells in same or other nodes. • Root – each node contains roots which are references to memory (cells) it considers useful. Ex – all cell references in the cpuregisters or in an executionstack are roots. • Reachable – cells referenced to by root directly or indirectly through other cells are live. • Unreachable – waste or unused memory

5. Terminology (contd) • Local reference – reference to cell (memory) on the same node or processor. • Remote reference – reference to cell on another node. • Entry & exit items – a remote reference to cell is represented by a reference to an exit item on the same node, which references an entry item on another node.

6. The basic Algorithm • Group Negotiation • Initial marking • Local propagation • Global propagation • Stabilization • Dead cycles removal • Group disbanding

7. GroupNegotiation • When a node decides to participate to new group GC, it first determines what group can be set up for that purpose. • Why? It could have been idle or its entry items may not have been accessed for a long time or it was not currently involved in any group GC. • How? Groups can be created based on geographic distances. Once created, a unique identifier is associated with each group GC and is made known to each of the nodes in the group.

8. Initial Marking • Within each group, entry items have a mark( soft or hard). • Hard – entry item referenced outside the group or roots. • Soft – referenced from within group. • Christopher’salgorithm used for marking – look at reference counter for entry item. Say it is K. If the number of exit items referencing that entry item within the group is K , then entry is marked soft. Else there is a reference from outside the group so mark it as Hard.

9. Local Propagation • Local GC are responsible for propagation of marks from entry items to exit items they reference locally, directly or indirectly. • 2 phase marking • Trace from entry items marked hard as well as from root set. Any exit item reached from this tracing is marked hard. • Start from soft entry items and mark all exit items reached by this tracing as soft

10. Global Propagation • Hard marks of exit items are propagated to their corresponding entry items within the group.

11. Stabilization • Group is said to be stable if • Nodes are stable, i.e they have no new data that could justify hardening more entry items locally or elsewhere in the group. • No new messages in transit that request the hardening of some entry item. • How? Group stability can be detected by any node termination detection algorithm.

12. Dead cycles removal • After stabilization, • Remove all entry items marked as soft.

13. Group disbanding • When a group GC is finished, its associated group may be disbanded. All data structures relative to this group can then be reclaimed.

14. Failure Handling • A node can detect failure of other nodes based on acks and time-outs. • A node that detect failure can • Decide that it is temporary and wait for failed node to wake up. • Re-organize the group i.e create a new group excluding the failed node.

15. Failure Handling(2) • A transmission link may fail and divide a group into subgroups. These subgroups start independent GC. (result – all dead memory will be cleaned)

16. Failure Handling(3) • When a node has a non-recoverable failure • What happens to entry referenced by failed node? Group G – calculate the number of entries and reference count of each entry. Suppose entry A has 4 references. Group G’ (G – failed node) do the same. Entry A now has 3 references. So A was being referenced by the failed node. Send a decrement message to that entry

17. Simultaneous Group Collections • A node may belong to more than one group. • Aim – The results obtained by a local garbage collector on a node can be used in other groups to which the node belongs. i.e markings can be used across the groups. • Adv – In a large n/w with variations in network connectivity and communication speed, GC is much more faster and efficient if groups are broken down into sub groups.

18. G G’ x x x x x x x x Group Contention(1) • Consider a subgroup G’ of G. If an entry item is marked hard in G then it can also be marked hard in G’. However some entry items that are soft w.r.t G will be hard w.r.t G’. Problem – Markings cannot be used across groups • GC of any one group can take place at a time.

19. G G’ x x x x x x x x Group Contention (2) • Conversely if a local Garbage Collector works for the group GC of G’, its hard marking cannot be used for G. However soft markings may be used.

20. Group Contention (3) • The situation is worse if a node belongs to overlapping groups. The markings (hard/soft) of local garbage collector with group A or B cannot be used at all for the other. A B

21. Contd… • This necessitates of having strictly hierarchical embedding of groups to avoid contention over the services of local garbage collector.

22. Hierarchical cooperation of group GCs • Soln to the group contention problem by modifying the marking scheme. • Groups are organized in a strictly hierarchical order by inclusion. • Each group is assigned a level index – number of groups it is strictly embedded in. • Universal group – level index 0

23. 1 2 3 H s Contd.. • In this scheme instead of binary hard and soft marks, we use integer marking scheme. REST OF ALGO REMAINS SAME. • Mark for a node entry is the least level for which the entry could be marked hard. • For group 3, left side is hard and right side is soft.

24. g0 g1 g2 X N1 g3 X x Marking scheme • Entry at N1 is hard w.r.t G2 • Entry at N1 is soft w.r.t G1 • - Marking of entry is 2

25. Contd.. • Local Propagation – Instead of prop H/S marks levels are prop from entry to exit items. • Rest of the algorithm is exactly SAME.

26. g0 g1 g2 X N1 g3 X x Final GC After prop of all marks, all Entry items With mark > 2, are soft w.r.t G2 hence Garbage collected.

27. Conclusion • Thus we saw a distributed GC algo which is • Fault tolerant. • Does not need a centralized control. • Allows for multiple concurrent active GCs. • Eventually reclaims all inaccessible objects including distributed cycles.