1. Cache Memory CSE 675.02 Slides from Dan Garcia, UCB

2. The Big Picture Processor (active) Computer Control (“brain”) Datapath (“brawn”) Memory (passive) (where programs, data live when running) Devices Input Output Keyboard, Mouse Display, Printer Disk, Network

3. Memory Hierarchy (1/3) Processor executes instructions on order of nanoseconds to picoseconds holds a small amount of code and data in registers Memory More capacity than registers, still limited Access time ~50-100 ns Disk HUGE capacity (virtually limitless) VERY slow: runs ~milliseconds

4. The Levels in Memory Hierarchy Higher the level, smaller and faster the memory. Try to keep most of the action in the higher levels.

5. Review: Why We Use Caches µProc 60%/yr. DRAM 7%/yr. 1 10 100 1000 19801981198319841985198619871988 1989 1990 1991199219931994199519961997199819992000 DRAM CPU 1982 Processor-Memory Performance Gap: (grows 50% / year) Performance “Moore’s Law” 1989 first Intel CPU with cache on chip 1998 Pentium III has two levels of cache on chip

6. Memory Hierarchy (2/3) Processor Size of memory at each level Increasing Distance from Proc., Decreasing speed Level 1 Level 2 Level n Level 3... Higher Lower Levels in memory hierarchy As we move to deeper levels the latency goes up and price per bit goes down. Q: Can $/bit go up as move deeper?

7. Memory Hierarchy (3/3) If level closer to Processor, it must be: smaller faster subset of lower levels (contains most recently used data) Lowest Level (usually disk) contains all available data Other levels?

8. Memory Caching We’ve discussed three levels in the hierarchy: processor, memory, disk Mismatch between processor and memory speeds leads us to add a new level: a memory cache Implemented with SRAM technology: faster but more expensive than DRAM memory. “S” = Static, no need to refresh, ~10ns “D” = Dynamic, need to refresh, ~60ns arstechnica.com/paedia/r/ram_guide/ram_guide.part1-1.html

9. Memory Hierarchy Analogy: Library (1/2) You’re writing a term paper (Processor) at a table in SEL SEL Library is equivalent to disk essentially limitless capacity very slow to retrieve a book Table is memory smaller capacity: means you must return book when table fills up easier and faster to find a book there once you’ve already retrieved it

10. Memory Hierarchy Analogy: Library (2/2) Open books on table are cache smaller capacity: can have very few open books fit on table; again, when table fills up, you must close a book much, much faster to retrieve data Illusion created: whole library open on the tabletop Keep as many recently used books open on table as possible since likely to use again Also keep as many books on table as possible, since faster than going to library

11. Memory Hierarchy Basis Disk contains everything. When Processor needs something, bring it into to all higher levels of memory. Cache contains copies of data in memory that are being used. Memory contains copies of data on disk that are being used. Entire idea is based on Temporal Locality: if we use it now, we’ll want to use it again soon (a Big Idea)

12. Cache Design How do we organize cache? Where does each memory address map to? (Remember that cache is subset of memory, so multiple memory addresses map to the same cache location.) How do we know which elements are in cache? How do we quickly locate them?

13. Direct-Mapped Cache (1/2) In a direct-mapped cache, each memory address is associated with one possible block within the cache Therefore, we only need to look in a single location in the cache for the data if it exists in the cache Block is the unit of transfer between cache and memory

14. Direct-Mapped Cache (2/2) Cache Location 0 can be occupied by data from: Memory location 0, 4, 8,... 4 blocks => any memory location that is multiple of 4 Memory Memory Address 0 1 2 3 4 5 6 7 8 9 A B C D E F 4 Byte Direct Mapped Cache Cache Index 0 1 2 3

15. Issues with Direct-Mapped Since multiple memory addresses map to same cache index, how do we tell which one is in there? What if we have a block size > 1 byte? Answer: divide memory address into three fields ttttttttttttttttt iiiiiiiiii oooo tagindexbyte to checkto offset if have selectwithin correct blockblockblock WIDTHHEIGHT Tag Index Offset

16. Direct-Mapped Cache Terminology All fields are read as unsigned integers. Index: specifies the cache index (which “row” of the cache we should look in) Offset: once we’ve found correct block, specifies which byte within the block we want -- I.e., which “column” Tag: the remaining bits after offset and index are determined; these are used to distinguish between all the memory addresses that map to the same location

17. TIO Dan’s great cache mnemonic AREA (cache size, B) = HEIGHT (# of blocks) * WIDTH (size of one block, B/block) WIDTH (size of one block, B/block) HEIGHT (# of blocks) AREA (cache size, B) 2 (H+W) = 2 H * 2 W Tag Index Offset

18. Direct-Mapped Cache Example (1/3) Suppose we have a 16KB of data in a direct-mapped cache with 4 word blocks Determine the size of the tag, index and offset fields if we’re using a 32-bit architecture Offset need to specify correct byte within a block block contains 4 words = 16 bytes = 2 4 bytes need 4 bits to specify correct byte

19. Direct-Mapped Cache Example (2/3) Index: (~index into an “array of blocks”) need to specify correct row in cache cache contains 16 KB = 2 14 bytes block contains 2 4 bytes (4 words) # blocks/cache =bytes/cache bytes/block = 2 14 bytes/cache 2 4 bytes/block =2 10 blocks/cache need 10 bits to specify this many rows

20. Direct-Mapped Cache Example (3/3) Tag: use remaining bits as tag tag length = addr length - offset - index = 32 - 4 - 10 bits = 18 bits so tag is leftmost 18 bits of memory address Why not full 32 bit address as tag? All bytes within block need same address (4b) Index must be same for every address within a block, so its redundant in tag check, thus can leave off to save memory (10 bits in this example)

21. Caching Terminology When we try to read memory, 3 things can happen: 1.cache hit: cache block is valid and contains proper address, so read desired word 2.cache miss: nothing in cache in appropriate block, so fetch from memory 3.cache miss, block replacement: wrong data is in cache at appropriate block, so discard it and fetch desired data from memory (cache always copy)

22. Accessing data in a direct mapped cache Ex.: 16KB of data, direct-mapped, 4 word blocks Read 4 addresses 1.0x00000014 2.0x0000001C 3.0x00000034 4.0x00008014 Memory values on right: only cache/ memory level of hierarchy Address (hex) Value of Word Memory 00000010 00000014 00000018 0000001C a b c d... 00000030 00000034 00000038 0000003C e f g h 00008010 00008014 00008018 0000801C i j k l...

23. Accessing data in a direct mapped cache 4 Addresses: 0x00000014, 0x0000001C, 0x00000034, 0x00008014 4 Addresses divided (for convenience) into Tag, Index, Byte Offset fields 000000000000000000 0000000001 0100 000000000000000000 0000000001 1100 000000000000000000 0000000011 0100 000000000000000010 0000000001 0100 Tag Index Offset

24. 16 KB Direct Mapped Cache, 16B blocks Valid bit: determines whether anything is stored in that row (when computer initially turned on, all entries invalid)... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... Index 0 0 0 0 0 0 0 0 0 0

25. 1. Read 0x00000014... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 000000000000000000 0000000001 0100 Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0 0 0

26. So we read block 1 (0000000001)... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 000000000000000000 0000000001 0100 Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0 0 0

27. No valid data... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 000000000000000000 0000000001 0100 Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0 0 0

28. So load that data into cache, setting tag, valid... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 1 0abcd 000000000000000000 0000000001 0100 Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0 0

29. Read from cache at offset, return word b 000000000000000000 0000000001 0100... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 1 0abcd Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0 0

30. 2. Read 0x0000001C = 0…00 0..001 1100... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 1 0abcd 000000000000000000 0000000001 1100 Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0 0

31. Index is Valid... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 1 0abcd 000000000000000000 0000000001 1100 Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0 0

32. Index valid, Tag Matches... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 1 0abcd 000000000000000000 0000000001 1100 Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0 0

33. Index Valid, Tag Matches, return d... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 1 0abcd 000000000000000000 0000000001 1100 Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0 0

34. 3. Read 0x00000034 = 0…00 0..011 0100... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 1 0abcd 000000000000000000 0000000011 0100 Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0 0

35. So read block 3... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 1 0abcd 000000000000000000 0000000011 0100 Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0 0

36. No valid data... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 1 0abcd 000000000000000000 0000000011 0100 Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0 0

37. Load that cache block, return word f... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 1 0abcd 000000000000000000 0000000011 0100 1 0efgh Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0

38. 4. Read 0x00008014 = 0…10 0..001 0100... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 1 0abcd 000000000000000010 0000000001 0100 1 0efgh Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0

39. So read Cache Block 1, Data is Valid... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 1 0abcd 000000000000000010 0000000001 0100 1 0efgh Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0

40. Cache Block 1 Tag does not match (0 != 2)... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 1 0abcd 000000000000000010 0000000001 0100 1 0efgh Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0

41. Miss, so replace block 1 with new data & tag... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 1 2ijkl 000000000000000010 0000000001 0100 1 0efgh Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0

42. And return word j... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7 1022 1023... 1 2ijkl 000000000000000010 0000000001 0100 1 0efgh Index Tag fieldIndex fieldOffset 0 0 0 0 0 0 0 0

43. Do an example yourself. What happens? Chose from: Cache: Hit, Miss, Miss w. replace Values returned:a,b, c, d, e,..., k, l Read address 0x00000030 ? 000000000000000000 0000000011 0000 Read address 0x0000001c ? 000000000000000000 0000000001 1100... Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3 4 5 6 7... 1 2ijkl 1 0efgh Index 0 0 0 0 0 0 Cache

44. Answers 0x00000030 a hit Index = 3, Tag matches, Offset = 0, value = e 0x0000001c a miss Index = 1, Tag mismatch, so replace from memory, Offset = 0xc, value = d Since reads, values must = memory values whether or not cached: 0x00000030 = e 0x0000001c = d Address Value of Word Memory 00000010 00000014 00000018 0000001c a b c d... 00000030 00000034 00000038 0000003c e f g h 00008010 00008014 00008018 0000801c i j k l...

45. Peer Instruction A. Mem hierarchies were invented before 1950. (UNIVAC I wasn’t delivered ‘til 1951) B. If you know your computer’s cache size, you can often make your code run faster. C. Memory hierarchies take advantage of spatial locality by keeping the most recent data items closer to the processor. ABC 1: FFF 2: FFT 3: FTF 4: FTT 5: TFF 6: TFT 7: TTF 8: TTT

46. Peer Instructions 1. All caches take advantage of spatial locality. 2. All caches take advantage of temporal locality. 3. On a read, the return value will depend on what is in the cache. ABC 1: FFF 2: FFT 3: FTF 4: FTT 5: TFF 6: TFT 7: TTF 8: TTT

47. And in Conclusion (1/2) We would like to have the capacity of disk at the speed of the processor: unfortunately this is not feasible. So we create a memory hierarchy: each successively lower level contains “most used” data from next higher level exploits temporal locality do the common case fast, worry less about the exceptions (design principle of MIPS) Locality of reference is a Big Idea

48. And in Conclusion (2/2) Mechanism for transparent movement of data among levels of a storage hierarchy set of address/value bindings address  index to set of candidates compare desired address with tag service hit or miss -load new block and binding on miss Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3... 1 0abcd 000000000000000000 0000000001 1100 address: tag index offset

49. Peer Instruction Answer A. Mem hierarchies were invented before 1950. (UNIVAC I wasn’t delivered ‘til 1951) B. If you know your computer’s cache size, you can often make your code run faster. C. Memory hierarchies take advantage of spatial locality by keeping the most recent data items closer to the processor. ABC 1: FFF 2: FFT 3: FTF 4: FTT 5: TFF 6: TFT 7: TTF 8: TTT A.“We are…forced to recognize the possibility of constructing a hierarchy of memories, each of which has greater capacity than the preceding but which is less accessible.” – von Neumann, 1946 B.Certainly! That’s call “tuning” C.“Most Recent” items  Temporal locality

50. Peer Instruction Answer 1. All caches take advantage of spatial locality. 2. All caches take advantage of temporal locality. 3. On a read, the return value will depend on what is in the cache. T R U E F A L S E 1. Block size = 1, no spatial! 2. That’s the idea of caches; We’ll need it again soon. 3. It better not! If it’s there, use it. Oth, get from mem F A L S E ABC 1: FFF 2: FFT 3: FTF 4: FTT 5: TFF 6: TFT 7: TTF 8: TTT

51. Caches Review… Mechanism for transparent movement of data among levels of a storage hierarchy set of address/value bindings address  index to set of candidates compare desired address with tag service hit or miss -load new block and binding on miss Valid Tag 0x0-3 0x4-70x8-b0xc-f 0 1 2 3... 1 0abcd 000000000000000000 0000000001 1100 address: tag index offset

52. Review: Why We Use Caches µProc 60%/yr. DRAM 7%/yr. 1 10 100 1000 19801981198319841985198619871988 1989 1990 1991199219931994199519961997199819992000 DRAM CPU 1982 Processor-Memory Performance Gap: (grows 50% / year) Performance “Moore’s Law” 1989 first Intel CPU with cache on chip 1998 Pentium III has two levels of cache on chip

53. Block Size Tradeoff (1/3) Benefits of Larger Block Size Spatial Locality: if we access a given word, we’re likely to access other nearby words soon Very applicable with Stored-Program Concept: if we execute a given instruction, it’s likely that we’ll execute the next few as well Works nicely in sequential array accesses too

54. Block Size Tradeoff (2/3) Drawbacks of Larger Block Size Larger block size means larger miss penalty -on a miss, takes longer time to load a new block from next level If block size is too big relative to cache size, then there are too few blocks -Result: miss rate goes up In general, minimize Average Memory Access Time (AMAT) = Hit Time + Miss Penalty x Miss Rate

55. Block Size Tradeoff (3/3) Hit Time = time to find and retrieve data from current level cache Miss Penalty = average time to retrieve data on a current level miss (includes the possibility of misses on successive levels of memory hierarchy) Hit Rate = % of requests that are found in current level cache Miss Rate = 1 - Hit Rate

56. Extreme Example: One Big Block Cache Size = 4 bytesBlock Size = 4 bytes Only ONE entry in the cache! If item accessed, likely accessed again soon But unlikely will be accessed again immediately! The next access will likely to be a miss again Continually loading data into the cache but discard data (force out) before use it again Nightmare for cache designer: Ping Pong Effect Cache Data Valid Bit B 0B 1B 3 Tag B 2

57. Block Size Tradeoff Conclusions Miss Penalty Block Size Increased Miss Penalty & Miss Rate Average Memory Access Time Block Size Exploits Spatial Locality Fewer blocks: compromises temporal locality Miss Rate Block Size

58. Types of Cache Misses (1/2) “Three Cs” Model of Misses 1st C: Compulsory Misses occur when a program is first started cache does not contain any of that program’s data yet, so misses are bound to occur can’t be avoided easily, so won’t focus on these in this course

59. Types of Cache Misses (2/2) 2nd C: Conflict Misses miss that occurs because two distinct memory addresses map to the same cache location two blocks (which happen to map to the same location) can keep overwriting each other big problem in direct-mapped caches how do we lessen the effect of these? Dealing with Conflict Misses Solution 1: Make the cache size bigger -Fails at some point Solution 2: Multiple distinct blocks can fit in the same cache Index?

60. Fully Associative Cache (1/3) Memory address fields: Tag: same as before Offset: same as before Index: non-existant What does this mean? no “rows”: any block can go anywhere in the cache must compare with all tags in entire cache to see if data is there

61. Fully Associative Cache (2/3) Fully Associative Cache (e.g., 32 B block) compare tags in parallel Byte Offset : Cache Data B 0 0 4 31 : Cache Tag (27 bits long) Valid : B 1B 31 : Cache Tag = = = = = :

62. Fully Associative Cache (3/3) Benefit of Fully Assoc Cache No Conflict Misses (since data can go anywhere) Drawbacks of Fully Assoc Cache Need hardware comparator for every single entry: if we have a 64KB of data in cache with 4B entries, we need 16K comparators: infeasible

63. Third Type of Cache Miss Capacity Misses miss that occurs because the cache has a limited size miss that would not occur if we increase the size of the cache sketchy definition, so just get the general idea This is the primary type of miss for Fully Associative caches.

64. N-Way Set Associative Cache (1/4) Memory address fields: Tag: same as before Offset: same as before Index: points us to the correct “row” (called a set in this case) So what’s the difference? each set contains multiple blocks once we’ve found correct set, must compare with all tags in that set to find our data

65. N-Way Set Associative Cache (2/4) Summary: cache is direct-mapped w/respect to sets each set is fully associative basically N direct-mapped caches working in parallel: each has its own valid bit and data

66. N-Way Set Associative Cache (3/4) Given memory address: Find correct set using Index value. Compare Tag with all Tag values in the determined set. If a match occurs, hit!, otherwise a miss. Finally, use the offset field as usual to find the desired data within the block.

67. N-Way Set Associative Cache (4/4) What’s so great about this? even a 2-way set assoc cache avoids a lot of conflict misses hardware cost isn’t that bad: only need N comparators In fact, for a cache with M blocks, it’s Direct-Mapped if it’s 1-way set assoc it’s Fully Assoc if it’s M-way set assoc so these two are just special cases of the more general set associative design

68. Associative Cache Example Recall this is how a simple direct mapped cache looked. This is also a 1-way set- associative cache! Memory Memory Address 0 1 2 3 4 5 6 7 8 9 A B C D E F 4 Byte Direct Mapped Cache Cache Index 0 1 2 3

69. Associative Cache Example Here’s a simple 2 way set associative cache. Memory Memory Address 0 1 2 3 4 5 6 7 8 9 A B C D E F Cache Index 0 0 1 1

70. Block Replacement Policy (1/2) Direct-Mapped Cache: index completely specifies which position a block can go in on a miss N-Way Set Assoc: index specifies a set, but block can occupy any position within the set on a miss Fully Associative: block can be written into any position Question: if we have the choice, where should we write an incoming block?

71. Block Replacement Policy (2/2) If there are any locations with valid bit off (empty), then usually write the new block into the first one. If all possible locations already have a valid block, we must pick a replacement policy: rule by which we determine which block gets “cached out” on a miss.

72. Block Replacement Policy: LRU LRU (Least Recently Used) Idea: cache out block which has been accessed (read or write) least recently Pro: temporal locality  recent past use implies likely future use: in fact, this is a very effective policy Con: with 2-way set assoc, easy to keep track (one LRU bit); with 4-way or greater, requires complicated hardware and much time to keep track of this

73. Block Replacement Example We have a 2-way set associative cache with a four word total capacity and one word blocks. We perform the following word accesses (ignore bytes for this problem): 0, 2, 0, 1, 4, 0, 2, 3, 5, 4 How many hits and how many misses will there be for the LRU block replacement policy?

74. Block Replacement Example: LRU Addresses 0, 2, 0, 1, 4, 0,... 0 lru 2 1 loc 0loc 1 set 0 set 1 02 lru set 0 set 1 0: miss, bring into set 0 (loc 0) 2: miss, bring into set 0 (loc 1) 0: hit 1: miss, bring into set 1 (loc 0) 4: miss, bring into set 0 (loc 1, replace 2) 0: hit 0 set 0 set 1 lru 02 set 0 set 1 lru set 0 set 1 0 1 lru 2 4 set 0 set 1 04 1 lru

75. Administrivia Do your reading! VM is coming up, and it’s shown to be hard for students! Any other announcements?

76. Big Idea How to choose between associativity, block size, replacement policy? Design against a performance model Minimize: Average Memory Access Time = Hit Time + Miss Penalty x Miss Rate influenced by technology & program behavior Note: Hit Time encompasses Hit Rate!!! Create the illusion of a memory that is large, cheap, and fast - on average

77. Example Assume Hit Time = 1 cycle Miss rate = 5% Miss penalty = 20 cycles Calculate AMAT… Avg mem access time = 1 + 0.05 x 20 = 1 + 1 cycles = 2 cycles

78. Ways to reduce miss rate Larger cache limited by cost and technology hit time of first level cache < cycle time More places in the cache to put each block of memory – associativity fully-associative -any block any line N-way set associated -N places for each block -direct map: N=1

79. Improving Miss Penalty When caches first became popular, Miss Penalty ~ 10 processor clock cycles Today 2400 MHz Processor (0.4 ns per clock cycle) and 80 ns to go to DRAM  200 processor clock cycles! Proc $2$2 DRAM $ MEM Solution: another cache between memory and the processor cache: Second Level (L2) Cache

80. Analyzing Multi-level cache hierarchy Proc $2$2 DRAM $ L1 hit time L1 Miss Rate L1 Miss Penalty Avg Mem Access Time = L1 Hit Time + L1 Miss Rate * L1 Miss Penalty L1 Miss Penalty = L2 Hit Time + L2 Miss Rate * L2 Miss Penalty Avg Mem Access Time = L1 Hit Time + L1 Miss Rate * (L2 Hit Time + L2 Miss Rate * L2 Miss Penalty ) L2 hit time L2 Miss Rate L2 Miss Penalty

81. Typical Scale L1 size: tens of KB hit time: complete in one clock cycle miss rates: 1-5% L2: size: hundreds of KB hit time: few clock cycles miss rates: 10-20% L2 miss rate is fraction of L1 misses that also miss in L2 why so high?

82. Example: with L2 cache Assume L1 Hit Time = 1 cycle L1 Miss rate = 5% L2 Hit Time = 5 cycles L2 Miss rate = 15% (% L1 misses that miss) L2 Miss Penalty = 200 cycles L1 miss penalty = 5 + 0.15 * 200 = 35 Avg mem access time = 1 + 0.05 x 35 = 2.75 cycles

83. Example: without L2 cache Assume L1 Hit Time = 1 cycle L1 Miss rate = 5% L1 Miss Penalty = 200 cycles Avg mem access time = 1 + 0.05 x 200 = 11 cycles 4x faster with L2 cache! (2.75 vs. 11)

84. CPU Cache Bus Memory One-Word-Wide Memory Bus Although caches are interested in low–latency main memory, it is generally easier to improve memory bandwidth with new organization than it is to reduce latency. Let us assume the performance of the main memory to be: – 1 clock cycle to send address, – 14 clock cycles for the access time per word, – 1 clock cycle to send word of data. Given a cache block of 1 word (1 word=4 bytes): Miss penalty = 1+14+1 = 16 cycles Throughput = 4 bytes / 16 cycles = ¼ bytes per cycle

85. CPU Cache Bus Memory a. One-word-wide memory organization CPU Bus b. Wide memory organization Memory Multiplexor Cache Wider Main Memory Bus CPU will still access cache a word at a time, so there is a need for a multiplexer. Given 4-word wide memory bus and a cache block of four words: Miss penalty 4-word memory bus = 1 + 14 + 1 = 16 cycles (as before) Throughput = 16 bytes / 16 cycles = 1 byte per cycle Figure 7.11 a. & b.

86. Wider Main Memory Bus & Level 2 Cache CPU Bus Memory Multiplexor Cache – But the multiplexer may be on the critical timing path. – Here, the second level cache can help since multiplexing can be between first- and second-level caches, not on critical timing path.

87. CPU Cache Bus Memory bank 1 Memory bank 2 Memory bank 3 Memory bank 0 Interleaved Memory Organization The example assumes word addressing. With byte addressing and 4 bytes per word, each of addresses would be multiple of 4. Memory bus width is 1 word. Assuming cache block of four words: Miss penalty = 1 + 14 + 4×1 = 19 cycles, Throughput = 16 bytes / 19 cycles = 0.84 bytes per cycle One address is sent to all banks, and each bank sends its data in its clock cycle. This is four-way interleaved memory.

88. What to do on a write hit? Write-through update the word in cache block and corresponding word in memory Write-back update word in cache block allow memory word to be “stale”  add ‘dirty’ bit to each block indicating that memory needs to be updated when block is replaced  OS flushes cache before I/O… Performance trade-offs?

89. Peer Instructions 1. In the last 10 years, the gap between the access time of DRAMs & the cycle time of processors has decreased. (I.e., is closing) 2. A 2-way set-associative cache can be outperformed by a direct-mapped cache. 3. Larger block size  lower miss rate ABC 1: FFF 2: FFT 3: FTF 4: FTT 5: TFF 6: TFT 7: TTF 8: TTT

90. Peer Instructions Answer 1. In the last 10 years, the gap between the access time of DRAMs & the cycle time of processors has decreased. (I.e., is closing) 2. A 2-way set-associative cache can be outperformed by a direct-mapped cache. 3. Larger block size  lower miss rate ABC 1: FFF 2: FFT 3: FTF 4: FTT 5: TFF 6: TFT 7: TTF 8: TTT 1. That was was one of the motivation for caches in the first place -- that the memory gap is big and widening. 2. Sure, consider the caches from the previous slides with the following workload: 0, 2, 0, 4, 2 2-way: 0m, 2m, 0h, 4m, 2m; DM: 0m, 2m, 0h, 4m, 2h 3. Larger block size  lower miss rate, true until a certain point, and then the ping-pong effect takes over

91. Generalized Caching We’ve discussed memory caching in detail. Caching in general shows up over and over in computer systems Filesystem cache Web page cache Game Theory databases / tablebases Software memoization Others? Big idea: if something is expensive but we want to do it repeatedly, do it once and cache the result.

92. An actual CPU -- Early PowerPC Cache 32 KiByte Instructions and 32 KiByte Data L1 caches External L2 Cache interface with integrated controller and cache tags, supports up to 1 MiByte external L2 cache Dual Memory Management Units (MMU) with Translation Lookaside Buffers (TLB) Pipelining Superscalar (3 inst/cycle) 6 execution units (2 integer and 1 double precision IEEE floating point)

93. Cache Things to Remember Caches are NOT mandatory: Processor performs arithmetic, memory stores data Caches simply make data transfers go faster Each Memory Hiererarchy level subset of next higher level Caches speed up due to temporal locality: store data used recently Block size > 1 wd spatial locality speedup: Store words next to the ones used recently Cache design choices: size of cache: speed v. capacity direct-mapped v. associative choice of N for N-way set assoc block replacement policy 2 nd level cache? 3 rd level cache? Write through v. write back? Use performance model to pick between choices, depending on programs, technology, budget,...

94. An Actual CPU – Pentium M

95. Peer Instruction 1. Increased associativity (1->2->4->8-way)  decreased or steady miss rate. 2. Increased associativity  increased cost & slower access time. 3. The ratio of costs of a “miss” vs. a “hit” are within an order of magnitude between VM & cache ABC 1: FFF 2: FFT 3: FTF 4: FTT 5: TFF 6: TFT 7: TTF 8: TTT