An Efficient Memory Management Technique That Improves Localities

Transcription

1 An Efficient Memory Management Technique That Improves Localities Krishna Kavi Mehran Rezaei Dept. of Electrical and Computer Engineering University of Alabama in Huntsville Ron K. Cytron Department of Computer Science Washington University K.Kavi 1

2 Outline Background and Related Work Allocation Policies New Memory Management Techniques Address Ordered Binary Tree (ABT) Segregated Binary Trees (SBT) Empirical Results Conclusion And Future Work Requirements Of Memory Management 1. Execution Efficiency Fast allocation and Deallocation 2. Fragmentation Efficiency Small internal and external fragmentation 3. Overhead Efficiency Small memory overheads And 4. Locality efficiency K.Kavi 2

3 Existing Memory Management Algorithms (Allocation Policies) Sequential Fit Algorithms First Fit, Next Fit, Best Fit Buddy System Binary Buddy, Double Buddy Segregated Free Lists Simple Segregated Storage, Segregated Fit Stephenson s Algorithm K.Kavi 3

4 Existing Memory Management Algorithms Sequential Fit Algorithms head size address tail available Chunk 16 bytes Memory overhead K.Kavi 4

5 Existing Memory Management Algorithms Sequential Fit Algorithms (First Fit) head size address tail available Chunk malloc (64 bytes) Malloc Request for 64 bytes) K.Kavi 5

6 Existing Memory Management Algorithms Buddy System 64 bytes request 20 is satisfied 128 bytes 256 bytes malloc(39 bytes) K.Kavi 6

7 calculate the index Existing Memory Management Algorithms Segregated Free Lists (Simple Segregated Storage) index size available chunks request is satisfied malloc(29 bytes) K.Kavi 7

8 Existing Memory Management Algorithms Stephenson s Algorithm head size s address available chunk s 2 = < s 1 s 3 = < s 1 20 malloc(70 bytes) s s bytes Memory overhead K.Kavi 8

9 Existing Memory Management Algorithms Stephenson s Algorithm head malloc (70 bytes) size s address s = < s available chunk s 2 s s 3 = < s malloc(70 bytes) K.Kavi 9

10 Stephenson s Algorithm head size s address available chunk s s 2 = < s 1 s 3 = < s 2 s K.Kavi 10

11 Existing Memory Management Algorithms Brief Comparison Storage Utilization Sequential Fits Segregated Free Lists Buddy System Execution Performance Buddy System Segregated Free Lists Sequential Fits K.Kavi 11

12 New Memory Management Algorithms Address Ordered Binary Tree(ABT) head size maximum size (left subtree) available chunk address maximum size (right subtree) 20 malloc(70 bytes) 21 free(object x) size 20 and address bytes Memory overhead K.Kavi 12

13 New Memory Management Algorithms ABT head size maximum size (left subtree) available chunk address maximum size (right subtree) 20 malloc(70 bytes) K.Kavi 13

14 New Memory Management Algorithms ABT head size maximum size (left subtree) available chunk address maximum size (right subtree) malloc (70 bytes) request is responded K.Kavi 14

15 New Memory Management Algorithms ABT head size maximum size (left subtree) = available chunk address maximum size (right subtree) free(object x) K.Kavi 15

16 New Memory Management Algorithms ABT head size maximum size (left subtree) available chunk address maximum size (right subtree) K.Kavi 16

17 New Memory Management Algorithms Segregated Binary Tree(SBT) <= 16 <= 32 <= 64 <= 128 <= 256 <= 512 > 512 K.Kavi 17

18 Empirical Results Benchmark Description check Exercises the error checking capabilities of JVM compress Modified Lempel Ziv db Database Manager jack PCCTS tool javac Java compiler used in JDK jess Expert System mpegaudio MPEG Layer 3 decompressor mtrt Ray tracer (Threaded version) raytrace Ray tracer K.Kavi 18

19 Empirical Results Table-1: Benchmark Statistics Total No. Total No. Total Bytes Average Average Benchmark Allocation Deallocaton of Memory Size of Size of Live Requests Requests Requested Requests Byte Memory Byte Simple 12,110 8, M M Check 46,666 41, M M Jess 81,858 74, M M Jack 80,815 73, M M MPEG 97,431 91, M M Average 63,776 57, M M K.Kavi 19

20 Empirical Results Storage Utilization Table 2: Exact Allocation Binary Tree Sequential Fit Bench Avg. Coale- Nodes Nodes Avg. Coale- Nodes Mark No. of scence Searched Searched No. scence Searched Name Nodes Frq At All. At De-All Nodes Frq. At All. Avg. Max Avg Max Avg. Max Simple Check Jess Jack MPEG Avg In average Binary Tree requires 146 nodes Vs nodes for Sequential Fit 146 * 28 = 4088 bytes (Binary Tree) Vs * 16 = bytes (Seq. Fit) Binary Tree searches 35 nodes during All. and 47 nodes during De-All. Sequential Fit searches 1843 during Allocation Performance K.Kavi 20

21 Empirical Results simple Jess MPEG Ave. Nodes Binary Tree Ave. Nodes Seq. Fits Nodes Searched Binary Tree Nodes searched Seq. Fits K.Kavi 21

22 Execution Performance Table 2: Ex act Allocat ion (All ocate d_siz e) Ğ (Req uested_ Siz e) = 0 Binar y T re e Imp le m enta ton Lin ke d Lis t Im p leme ntat ion Ben c h - Av era ge C oa le - Node s S ear ch ed Node s S ear ch ed Av era ge C oa le - Node s S ear ch ed Ma rk # No de s sc en ce At Allo ca tion At Deallocation # No de s sc en se At Allo ca tion Nam e Frq Avg. Max Avg Max Frq Avg. max Si mp le C he ck J e ss J a ck MPE G Av g K.Kavi 22

23 Execution Performance Segregated Binary Tree check compress db jack javac jess mpegaudio mtrt raytrace Average Average Number of Nodes Simple Segregated Storage K.Kavi 23

24 Cache Performance Our Approach Collect address traces from Benchmark programs Replace object address use addresses based on the various memory allocation/de-allocation methods Use Cache simulators (Dinero) to collect cache performance data K.Kavi 24

25 Cache Performance Binary Tree Best Fit Binary Tree First Fit unix 2.5E+6 64 k Linked List Best Fit Linked List First Fit 2.0E+6 1.5E+6 1.0E E E Block Size(bytes) Miss rates for Check (64K cache) K.Kavi 25

26 Cache Performance bfbt bfdll ffbt ffdll unix 1.6E+6 1.4E+6 1.2E+6 1.0E E E E E E Block Size (bytes) Miss rates for Check (256K cache) K.Kavi 26

27 Cache Performance bfbt bfdll ffbt ffdll unix fib check jess jack mpeg Benchmark Miss rates for 4 benchmarks (64k, 16byte blocks) K.Kavi 27

28 Cache Performance Exact vs Over Allocation Do not keep too small chunks May increase internal fragementation Table 3: 16 byte Overallocation Binary Tree Implementaton Sequential Fit Bench- Mark Name Avg. Avg. Coale- Nodes Searched Nodes Searched Avg. Avg. Coale- Nodes Searched No. Intern scence At Allocation At De-allocation No. Intern scense At Allocation Nodes Frag- Frq Avg. Max Avg. Max Nodes Frag- Frq Avg. Max ment ment Simple Check Jess Jack MPEG Avg K.Kavi 28

29 Cache Performance Effect Of Over-Allocation Block Size Effect of Over-allocation for Check (256K cache) K.Kavi 29

30 Conclusions Binary Tree Consistently outperforms Seq. Fits Memory Overhead Nodes Searched at Allocation Fragmentation and Coalescence frequency Binary tree also performs better in terms of cache misses We have been comparing ABT and SBT with simple segregated list methods K.Kavi 30

31 Conclusions ABT, BF, 0 Overall SBT, BF, 16 Overall SBT, BF, 0 Overall SSS Average Number of Nodes (Y axis log scale) K.Kavi 31

32 Future Work Off load all memory management functions to IRAM devices Allocation - Deallocation Garbage Collection Improving Cache performance using Memory Forwarding Perform memory forwarding inside IRAM Transparently of CPU Improving pre-fetching performance using Jump Pointers Perform jump pointers inside IRAM K.Kavi 32

33 Memory Management Loading a Process into the Memory Header Info. Code Static Data Code Executable Code Dynamic Data Static Data Initialized Data Uninit. Data Load Module on Disk Heap Growth Direction Unused Logical Address Space Stack Stack Growth Direction Dynamic Data Stack Process Physical Address Space Process Logical Address Space K.Kavi 33

34 Memory Management Memory Management Hierarchy Operating System OS Memory Manager sbrk Free Chunk Page Size Process Memory Manager Process Memory Manager... Process Memory Manager Process 1 Process 2 Process N K.Kavi 34