Efficient and Reliable Lock-Free Memory Reclamation Based on Reference Counting

Transcription

1 Efficient and Reliable Lock-Free Memory Reclamation d on Reference ounting nders Gidenstam, Marina Papatriantafilou, Håkan Sundell and Philippas Tsigas Distributed omputing and Systems group, Department of omputer Science and Engineering, halmers University of Technology

2 Outline Introduction The Problem Lock-free synchronization Our solution Idea Properties Experiments onclusions 2

3 The Lock-Free Memory Reclamation Problem oncurrent shared data structure Dynamic use of shared memory oncurrent and overlapping operations by threads or processes 3

4 The Lock-Free Memory Reclamation Problem Local variables 4

5 The Lock-Free Memory Reclamation Problem X has de-referenced the link (pointer) to 5

6 The Lock-Free Memory Reclamation Problem nother thread, Y, finds and deletes (removes) from the active structure Thread Y 6

7 The Lock-Free Memory Reclamation Problem Property I: (de-)referenced node is not reclaimed Thread Y wants to reclaim(/free)? Thread Y 7

8 The Lock-Free Memory Reclamation Problem Property II: Links in a (de-)referenced node should always be de-referencable. The nodes and? are deleted from the active structure. D 8

9 The Lock-Free Memory Reclamation Problem Solutions? Garbage collection? Reference counting? Needs to be lock-free! 2 D 9

10 Lock-free synchronization lock-free shared data structure llows concurrent operations without enforcing mutual exclusion (i.e. no locks) Guarantees that at least one operation always makes progress voids: locking, deadlock and priority inversion Hardware synchronization primitives uilt into PU and memory system Typically: atomic read-modify-write instructions Examples Test-and-set, ompare-and-swap, Load-Linked / Store-onditional 0

11 Previous solutions Lock-free Reference ounting Valois + Michael & Scott 995 Detlefs et al. 200 Slow Herlihy et al Remaining issues slow thread might prevent reclamation yclic garbage Implementation practicality issues Reference-count field MUST remain forever (Valois + Michael & Scott) Needs double word S (Detlefs et al.) Needs double width S (Herlihy, 2002) Large overhead

12 Our approach The basic idea ombine the best of Hazard pointers (Michael 2002) Tracks references from threads Fast de-reference Upper bound on the amount of unreclaimed deleted nodes ompatible with standard memory allocators Reference counting Tracks references from links in shared memory Manages links within dynamic nodes Safe to traverse links (also) in deleted nodes Practical Uses only single-word ompare-nd-swap 2

13 The basic idea PI DeRefLink ReleaseRef omparendswapref StoreRef NewNode DeleteNode Hazard pointers () Deletion list 3

14 The basic idea PI DeRefLink() ReleaseRef omparendswapref StoreRef NewNode DeleteNode Hazard pointers () Deletion list R 4

15 The basic idea PI DeRefLink ReleaseRef(R) omparendswapref StoreRef NewNode DeleteNode Hazard pointers () Deletion list R 5

16 The basic idea PI DeRefLink ReleaseRef omparendswapref StoreRef NewNode 0 DeleteNode D Hazard pointers () Deletion list new 6

17 The basic idea PI DeRefLink ReleaseRef omparendswapref StoreRef(new.next, R) NewNode 0 DeleteNode D Hazard pointers () Deletion list R new 2 7

18 The basic idea PI DeRefLink ReleaseRef Hazard pointers () omparendswapref(prev.next, old, new) StoreRef NewNode DeleteNode D Deletion list prev old new 0 2 8

19 The basic idea PI DeRefLink ReleaseRef omparendswapref StoreRef NewNode DeleteNode(old) D Hazard pointers () Deletion list _ prev old new 0 2 9

20 reaking chains of garbage lean-up deleted nodes Update links to point to live nodes Performed on nodes in Own deletion list ll deletion lists 0 Hazard pointers (Thread Y) Deletion list 2 D E 20

21 reaking chains of garbage lean-up deleted nodes Update links to point to live nodes Performed on nodes in Own deletion list ll deletion lists 0 Hazard pointers (Thread Y) 0 Deletion list 3 D E 2

22 ound on unreclaimed nodes deleted node can be reclaimed when The reference count is zero and No hazard pointer is pointing to it and There is no ongoing clean-up of this node With a rate relative to the number of threads of Scanning hazard pointers leaning up nodes as needed Then the maximum size of each deletion list depends on The number of hazard pointers The number of links per node The number of threads 22

23 Experimental evaluation Lock-free deque (Sundell and Tsigas 2004) (deque double-ended queue) The algorithm needs traversal of deleted nodes Time for 0000 random operations/thread Tested memory reclamation schemes Reference counting, Valois et al. The new algorithm Systems 4 processor Xeon P / Linux (UM) 8 processor SGI Origin 2000 / IRIX (NUM) 23

24 Experimental evaluation 24

25 Experimental evaluation 25

26 onclusions First lock-free memory reclamation scheme that Only uses atomic primitives available in contemporary architectures Guarantees safety of Local and Global references Has an upper bound on the amount of deleted but unreclaimed nodes llows arbitrary reuse of reclaimed memory 26

27 Questions? ontact Information: ddress: nders Gidenstam, omputer Science & Engineering, halmers University of Technology, SE Göteborg, Sweden cs.chalmers.se Web: Implementation 27

28 onclusions First lock-free memory reclamation scheme that Only uses atomic primitives available in contemporary architectures Guarantees safety of Local and Global references Has an upper bound on the amount of deleted but unreclaimed nodes ( ound: N * N * (k + L_max + a + ) ) llows arbitrary reuse of reclaimed memory 28