3parallel memoryallocation
Download as PPT, PDF2 likes521 views
The document discusses parallel memory allocation strategies for multithreaded applications. It defines key terms like heap and dynamic memory allocator. It examines various strategies for parallel allocation like using a concurrent single heap with multiple free lists, multiple heaps, private heaps, or a combination of global and local heaps. Specific allocators discussed include Buddy System, MFLF, NUMAmalloc, Hoard, and VH. The performance tradeoffs between speed, scalability, fragmentation, and false sharing are also summarized.
3parallel memoryallocation
- 1. Parallel Memory AllocationParallel Memory Allocation Steven Saunders
- 2. Steven Saunders 2 Parallel Memory Allocation Introduction Fallacy: All dynamic memory allocators are either scalable, effective or efficient... Truth: no allocator exists that handles all situations, allocation patterns and memory hierarchies best (serial or parallel) Research results are difficult to compare – no standard benchmarks, simulators vs. real tests Distributed memory, garbage collection, locality not considered in this talk
- 3. Steven Saunders 3 Parallel Memory Allocation Definitions Heap – pool of memory available for allocation or deallocation of arbitrarily-sized blocks in arbitrary order that will live an arbitrary amount of time Dynamic Memory Allocator – used to request or return memory blocks to the heap – aware only of the size of a memory block, not its type and value – tracks which parts of the heap are in use and which parts are available for allocation
- 4. Steven Saunders 4 Parallel Memory Allocation Design of an Allocator Strategy - consider regularities in program behavior and memory requests to determine a set of acceptable policies Policy - decide where to allocate a memory block within the heap Mechanism - implement the policy using a set of data structures and algorithms Emphasis has been on policies and mechanisms!
- 5. Steven Saunders 5 Parallel Memory Allocation Strategy Ideal Serial Strategy – “put memory blocks where they won’t cause fragmentation later” – Serial program behavior: ramps, peaks, plateaus Parallel Strategies – “minimize unrelated objects on the same page” – “bound memory blowup and minimize false sharing” – Parallel program behavior: SPMD, producer-consumer
- 6. Steven Saunders 6 Parallel Memory Allocation Policy Common Serial Policies: – best fit, first fit, worst fit, etc. Common Techniques: – splitting - break large blocks into smaller pieces to satisfy smaller requests – coalescing - free blocks to satisfy bigger requests immediate - upon deallocation deferred - wait until requests cannot be satisfied
- 7. Steven Saunders 7 Parallel Memory Allocation Mechanism Each block contains header information – size, in use flag, pointer to next free block, etc. Free List - list of memory blocks available for allocation – Singly/Doubly Linked List - each free block points to the next free block – Boundary Tag - size info at both ends of block positive indicates free, negative indicates in use – Quick Lists - multiple linked lists, where each list contains blocks of equal size
- 8. Steven Saunders 8 Parallel Memory Allocation Performance Speed - comparable to serial allocators Scalability - scale linearly with processors False Sharing - avoid actively causing it Fragmentation - keep it to a minimum
- 9. Steven Saunders 9 Parallel Memory Allocation 0 1 Memory Cache Performance (False Sharing) Multiple processors inadvertently share data that happens to reside in the same cache line – padding solves problem, but greatly increases fragmentation 0 1 0 1 0 1 X X X X
- 10. Steven Saunders 10 Parallel Memory Allocation Performance (Fragmentation) Inability to use available memory External - available blocks cannot satisfy requests (e.g., too small) Internal - block used to satisfy a request is larger than the requested size
- 11. Steven Saunders 11 Parallel Memory Allocation Performance (Blowup) Out of control fragmentation – unique to parallel allocators – allocator correctly reclaims freed memory but fails to reuse it for future requests available memory not seen by allocating processor – p threads that serialize one after another require O(s) memory – p threads that execute in parallel require O(ps) memory – p threads that execute interleaved on 1 processor require O(ps) memory... ... x = malloc(s); free(x); ... Example: blowup from exchanging memory (Producer-Consumer)
- 12. Steven Saunders 12 Parallel Memory Allocation Parallel Allocator Taxonomy Serial Single Heap – global lock protects the heap Concurrent Single Heap – multiple free lists or a concurrent free list Multiple Heaps – processors allocate from any of several heaps Private Heaps – processors allocate exclusively from a local heap Global & Local Heaps – processors allocate from a local and a global heap
- 13. Steven Saunders 13 Parallel Memory Allocation Serial Single Heap Make an existing serial allocator thread-safe – utilize a single global lock for every request Performance – high speed, assuming a fast lock – scalability limited – false sharing not considered – fragmentation bounded by serial policy Typically production allocators – IRIX, Solaris, Windows
- 14. Steven Saunders 14 Parallel Memory Allocation Concurrent Single Heap Apply concurrency to existing serial allocators – use quick list mechanism, with a lock per list Performance – moderate speed, could require many locks – scalability limited by number of requested sizes – false sharing not considered – fragmentation bounded by serial policy Typically research allocators – Buddy System, MFLF (multiple free list first), NUMAmalloc
- 15. Steven Saunders 15 Parallel Memory Allocation Concurrent Single (Buddy System) Policy/Mechanism – one free list per memory block size: 1,2,4,…,2i – blocks recursively split on list i into 2 buddies for list i-1 in order to satisfy smaller requests – only buddies are coalesced to satisfy larger requests – each free list can be individually locked – trade speed for reduced fragmentation if free list empty, a thread’s malloc enters a wait queue malloc could be satisfied by another thread freeing memory or by breaking a higher list’s block into buddies (whichever finishes first) reducing the number of splits reduces fragmentation by leaving more large blocks for future requests
- 16. Steven Saunders 16 Parallel Memory Allocation Concurrent Single (Buddy System) Performance – moderate speed, complicated locking/queueing, although buddy split/coalesce code is fast – scalability limited by number of requested sizes – false sharing very likely – high internal fragmentation 1 2 4 8 16 Thread 1 Thread 2 x = malloc(8); 1 x = malloc(5); 2 y = malloc(8); free(x); 1 2 4 8 16 1
- 17. Steven Saunders 17 Parallel Memory Allocation Concurrent Single (MFLF) Policy/Mechanism – set of quick lists to satisfy small requests exactly malloc takes first block in appropriate list free returns block to head of appropriate list – set of misc lists to satisfy large requests quickly each list labeled with range of block sizes, low…high malloc takes first block in list where request < low – trades linear search for internal fragmentation free returns blocks to list where low < request < high – each list can be individually locked and searched
- 18. Steven Saunders 18 Parallel Memory Allocation Concurrent Single (MFLF) Performance – high speed – scalability limited by number of requested sizes – false sharing very likely – high internal fragmentation Approach similar to current state-of-the-art serial allocators
- 19. Steven Saunders 19 Parallel Memory Allocation Concurrent Single (NUMAmalloc) Strategy - minimize co-location of unrelated objects on the same page – avoid page level false sharing (DSM/software DSM) Policy - place same-sized requests in the same page (heuristic hypothesis) – basically MFLF on the page level Performance – high speed – scalability limited by number of requested sizes – false sharing: helps page level but not cache level – high internal fragmentation
- 20. Steven Saunders 20 Parallel Memory Allocation Multiple Heaps List of multiple heaps – individually growable, shrinkable and lockable – threads scan list looking for first available (trylock) – threads may cache result to reduce next lookup Performance – moderate speed, limited by # list scans and lock – scalability limited by number of heaps and traffic – false sharing unintentionally reduced – blowup increased (up to O(p)) Typically production allocators – ptmalloc (Linux), HP-UX
- 21. Steven Saunders 21 Parallel Memory Allocation Private Heaps Processors exclusively utilize a local private heap for all allocation and deallocation – eliminates need for locks Performance – extremely high speed – scalability unbounded – reduced false sharing pass memory to another thread – blowup unbounded Both research and production allocators – CILK, STL
- 22. Steven Saunders 22 Parallel Memory Allocation Global & Local Heaps Processors generally utilize a local heap – reduces most lock contention – private memory is acquired/returned to global heap (which is always locked) as necessary Performance – high speed, less lock contention – scalability limited by number of locks – low false sharing – blowup bounded (O(1)) Typically research allocators – VH, Hoard
- 23. Steven Saunders 23 Parallel Memory Allocation Global & Local Heaps (VH) Strategy - exchange memory overhead for improved scalability Policy – memory broken into stacks of size m/2 – global heap maintains a LIFO of stack pointers that local heaps can use global push releases a local heap’s stack global pop acquires a local heap’s stack – local heaps maintain an Active and Backup stack local operations private, i.e. don’t require locks Mechanism - serial free list within stacks
- 24. Steven Saunders 24 Parallel Memory Allocation Global & Local Heaps (VH) Memory Usage = M + m*p – M = amount of memory in use by program – p = number of processors – m = private memory (2 size m/2 stacks) higher m reduces number of global heap lock operations lower m reduces memory usage overhead B 0 A B 1 A B p A global heap private heap
- 25. Steven Saunders 25 Parallel Memory Allocation Global & Local Heaps (Hoard) Strategy - bound blowup and min. false sharing Policy – memory broken into superblocks of size S all blocks within a superblock are of equal size – global heap maintains a set of available superblocks – local heaps maintain local superblocks malloc satisfied by local superblock free returns memory to original allocating superblock (lock!) superblocks acquired as necessary from global heap if local usage drops below a threshold, superblocks are returned to the global heap Mechanism – private superblock quick lists
- 26. Steven Saunders 26 Parallel Memory Allocation Global & Local Heaps (Hoard) Memory Usage = O(M + p) – M = amount of memory in use by program – p = number of processors False Sharing – since malloc is satisfied by a local superblock, and free returns memory to the original superblock, false sharing is greatly reduced – worst case: a non-empty superblock is released to global heap and another thread acquires it set superblock size and emptiness threshold to minimize
- 27. Steven Saunders 27 Parallel Memory Allocation Global & Local Heaps (Compare) VH proves a tighter memory bound than Hoard and only requires global locks Hoard has a more flexible local mechanism and considers false sharing They’ve never been compared! – Hoard is production quality and would likely win
- 28. Steven Saunders 28 Parallel Memory Allocation Summary Memory allocation is still an open problem – strategies addressing program behavior still uncommon Performance Tradeoffs – speed, scalability, false sharing, fragmentation Current Taxonomy – serial single heap, concurrent single heap, multiple heaps, private heaps, global & local heaps
- 29. Steven Saunders 29 Parallel Memory Allocation References Serial Allocation 1) Paul Wilson, Mark Johnstone, Michael Neely, David Boles. Dynamic Storage Allocation: A Survey and Critical Review. 1995 International Workshop on Memory Management. Shared Memory Multiprocessor Allocation Concurrent Single Heap 2) Arun Iyengar. Scalability of Dynamic Storage Allocation Algorithms. Sixth Symposium on the Frontiers of Massively Parallel Computing. October 1996. 3) Theodore Johnson, Tim Davis. Space Efficient Parallel Buddy Memory Management. The Fourth International Conference on Computing and Information (ICCI'92). May 1992. 4) Jong Woo Lee, Yookun Cho. An Effective Shared Memory Allocator for Reducing False Sharing in NUMA Multiprocessors. IEEE Second International Conference on Algorithms & Architectures for Parallel Processing (ICAPP'96). Multiple Heaps 5) Wolfram Gloger. Dynamic Memory Allocator Implementations in Linux System Libraries. Global & Local Heaps 6) Emery Berger, Kathryn McKinley, Robert Blumofe, Paul Wilson. Hoard: A Scalable Memory Allocator for Multithreaded Applications. The Ninth International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS-IX). November 2000. 7) Voon-Yee Vee, Wen-Jing Hsu. A Scalable and Efficient Storage Allocator on Shared-Memory Multiprocessors. The International Symposium on Parallel Architectures, Algorithms, and Networks (I-SPAN'99). June 1999.