1. 1  1998 Morgan Kaufmann Publishers Chapter Seven Large and Fast: Exploiting Memory Hierarchy

2. 2  1998 Morgan Kaufmann Publishers SRAM: –value is stored on a pair of inverting gates –very fast but takes up more space than DRAM (4 to 6 transistors) DRAM: –value is stored as a charge on capacitor (must be refreshed) –very small but slower than SRAM (factor of 5 to 10) Memories: Review

3. 3  1998 Morgan Kaufmann Publishers Users want large and fast memories! SRAM access times are 2 - 25ns at cost of $100 to $250 per Mbyte. DRAM access times are 60-120ns at cost of $5 to $10 per Mbyte. Disk access times are 10 to 20 million ns at cost of $.10 to $.20 per Mbyte. Try and give it to them anyway –build a memory hierarchy Exploiting Memory Hierarchy 1997 CPU Level n Level 2 Level 1 Levels in the memory hierarchy Increasing distance from the CPU in access time Size of the memory at each level

4. 4  1998 Morgan Kaufmann Publishers Locality A principle that makes having a memory hierarchy a good idea If an item is referenced, temporal locality: it will tend to be referenced again soon spatial locality: nearby items will tend to be referenced soon. Why does code have locality? Our initial focus: two levels (upper, lower) –block: minimum unit of data –hit: data requested is in the upper level –miss: data requested is not in the upper level

5. 5  1998 Morgan Kaufmann Publishers Two issues: –How do we know if a data item is in the cache? –If it is, how do we find it? Our first example: – block size is one word of data – "direct mapped" For each item of data at the lower level, there is exactly one location in the cache where it might be. e.g., lots of items at the lower level share locations in the upper level Cache

6. 6  1998 Morgan Kaufmann Publishers Mapping: address is modulo the number of blocks in the cache Direct Mapped Cache

7. 7  1998 Morgan Kaufmann Publishers Example Memory: 32 words ==> 5 bits Cache: 8 words ==> 3 bits 00001, 01001, 10001, 11001 in memory maps to cache 001 (use the lower log2(8) bits, equivalent to mod 8) Similarly, 00101, 01101, 10101, 11101 maps to cache 101

8. 8  1998 Morgan Kaufmann Publishers Tags The mapping from memory location to cache location is straightforward (many-to-one mapping), but how about the other way around (one-to-many mapping)? In other words, how do we know whether the data in the cache corresponds to a requested word? Need to add a set of tags to the cache. These tags contain the address information required to identify whether a hit occurs. Tags contain the information we throw away in the many-to-one mapping, i.e., the upper portion of the address. For example, we need 2-bit tags in Figure 7.5. Also need an additional bit, called the valid bit, to indicate whether a cache block has valid information. (cache is empty at start up) Figure 7.6 illustrates the contents of an 8-word direct-mapped cache as it responds to a series of requests from the processor.

9. 9  1998 Morgan Kaufmann Publishers For MIPS: What kind of locality are we taking advantage of? Direct Mapped Cache Address (showing bit positions) 20 10 Byte offset ValidTagDataIndex 0 1 2 1021 1022 1023 Tag Index HitData 20 32 31 30 13 12 11 2 1 0

10. 10  1998 Morgan Kaufmann Publishers Cache Size Each unit: Block size + tag size + valid field size Assuming 32-bit byte address, a direct-mapped cache of size 2^n words with one-word (4-byte) block will require: 2^n x ( 32 + (32 - n - 2) + 1) = 2^n x (63-n) Example: How many total bits are required for a direct-mapped cache with 64KB of data and one-word blocks, assuming 32-bit address? 64 KB = 16 K words = 2^14 words = 2^14 blocks Each block has 32 bits of data plus a tag, which is 32-14-2 bits, plus a valid bit. Total cache size = 2^14 x (32 +(32 -14 -2) +1 ) = 2^14x 49 = 784x 2^10 = 784 Kbits = 96 KB.

11. 11  1998 Morgan Kaufmann Publishers Read hits –this is what we want! Read misses –stall the CPU, fetch block from memory, deliver to cache, restart –different from pipelining since we must continue executing some instructions while stalling others there. Write hits: –can replace data in cache and memory (write-through) –write the data to the cache and the write buffer (write-buffer) –write the data only into the cache (write-back the cache later) Write misses: –read the entire block into the cache, then write the word Hits vs. Misses

12. 12  1998 Morgan Kaufmann Publishers Dealing with instruction cache miss Send the original PC value ( current PC -4 ) to the memory Instruct the main memory to perform a read and wait for the memory to complete its access. Write the cache entry. Restart the instruction execution at the first step, which will refetch the instruction, this time finding it in the cache. What about data cache miss?

13. 13  1998 Morgan Kaufmann Publishers DECStation 3100 Cache One instruction cache, one data cache Each cache 64 KB, or 16K words, with a one-word block Use write-through scheme, writing the data into both the memory and the cache. Address (showing bit positions) 1614 Byte offset ValidTagData HitData 1632 16K entries 16 bits32 bits 31 30 17 16 15 5 4 3 2 1 0

14. 14  1998 Morgan Kaufmann Publishers Taking advantage of spatial locality: Cache block = (Block address) mod (Number of Cache Blocks) Direct Mapped Cache

15. 15  1998 Morgan Kaufmann Publishers Mapping an Address to a Multiword Cache Block Consider a cache with 64 blocks and a block size of 16 bytes. What block number does byte address 1200 map to?

16. 16  1998 Morgan Kaufmann Publishers Miss Rate versus Block Size 1 KB 8 KB 16 KB 64 KB 256 KB 256 40% 35% 30% 25% 20% 15% 10% 5% 0% M i s s r a t e 64164 Block size (bytes)

17. 17  1998 Morgan Kaufmann Publishers Make reading multiple words easier by using banks of memory It can get a lot more complicated... Hardware Issues

18. 18  1998 Morgan Kaufmann Publishers Increasing the block size tends to decrease miss rate: Use split caches because there is more spatial locality in code: Performance 1 KB 8 KB 16 KB 64 KB 256 KB 256 40% 35% 30% 25% 20% 15% 10% 5% 0% M i s s r a t e 64164 Block size (bytes)

19. 19  1998 Morgan Kaufmann Publishers Performance Simplified model: execution time = (execution cycles + stall cycles) x cycle time stall cycles = # of instructions x miss ratio x miss penalty Two ways of improving performance: –decreasing the miss ratio –decreasing the miss penalty What happens if we increase block size?

20. 20  1998 Morgan Kaufmann Publishers Calculating Cache Performance Assume an instruction cache miss rate for gcc of 2% and a data cache miss rate of 4%. If a machine has a CPI of 2 without any memory stalls and the miss penalty is 40 cycles for all misses, determined how much faster a machine would run with a perfect cache that never missed. (Use instruction frequencies for gcc from fig. 4.54)

21. 21  1998 Morgan Kaufmann Publishers Cache Performance with Increased Clock Rate Relative cache penalties increase as machine becomes faster!

22. 22  1998 Morgan Kaufmann Publishers Direct-mapped vs Fully-Associative Scheme In a direct-mapped scheme, there is a direct mapping from any block address in memory to a single location in the upper level of the hierarchy. There exist other possibilities. For example, a block can be placed in any location in the cache. This is called the fully associative scheme. In a fully associative scheme, we need to search all in entries in the cache to find a given block. Additional hardware (comparator) is required. Thus fully associative scheme is practical only for caches with small number of blocks. Between direct-mapped and fully-associative is the set-associative scheme.

23. 23  1998 Morgan Kaufmann Publishers Set-Associative Scheme In a set-associative cache, there are a fixed number of locations (at least 2) where each block can be placed. A set-associative cache with n locations for a block is called an n- way associative cache. An n-way associative cache consists of a number of sets, each of which consists of n blocks. The set containing the memory block: (Block number) modulo (Number of sets in the cache) We then need to perform a search to find the block in the set.

24. 24  1998 Morgan Kaufmann Publishers Location of Memory Block -- Example Given a memory block whose address is 12:

25. 25  1998 Morgan Kaufmann Publishers Decreasing miss ratio with associativity Note, however, that hit time is increased with increasing degree of associativity.

26. 26  1998 Morgan Kaufmann Publishers Locating a Block in the Cache Each block in a set-associative cache includes an address tag that gives the block address. The index value is used to select the set containing the address of interest. Block offset is the address of the desired data within the block. What is the value of block offset in a direct-mapped cache? TagIndexBlock Offset

27. 27  1998 Morgan Kaufmann Publishers Size of Tags vs Set Associativity If the total cache size is kept the same, increasing the number of associativity increases the number of blocks per set, which is the number of parallel comparison required. Each increase by a factor of two in associativity doubles the number of blocks per set and halves the number of sets. Thus each increase by a factor of two in associativity decreases the size of the index by 1 and increases the size of the tag by 1.

28. 28  1998 Morgan Kaufmann Publishers An implementation (4-way association)

29. 29  1998 Morgan Kaufmann Publishers Example Assume a cache of 4K blocks and a 32 bit address 4K = 2^12 For a direct-mapped scheme, we need (32-12) bits for a tag. Total # of tags = 4K. Thus total number of tag bits = 20x4K = 80Kbits. For a 2-way associative scheme, we need (32- 11) bits for a tag. Total # of tags = 2 x 2K = 4K, thus total number of tag bits = (32- 11)x4K = 84 Kbits. For a 4-way associative scheme, we need (32- 10) bits for a tag. Total # of tags = 4 x 1K = 4K, thus total number of tag bits = (32- 10)x4K = 88 Kbits. For a fully associative scheme, we need 32 bits for a tag. Total # of tags = 4K x 1 = 4K, thus total number of tag bits = 32x4K = 128 Kbits. The choice among direct-mapped, set-associative or fully associative mapping depends on the cost of a miss versus the cost of implementing associativity, both in time and in extra hardware.

30. 30  1998 Morgan Kaufmann Publishers Choosing Which Block to Replace In an associative cache, we have a choice of where to place the requested block, an hence a choice of which block to replace when a miss occurs. The most commonly used scheme is least recently used (LRU). The block replaced is the one that has been unused for the longest time. Again, extra hardware is needed to keep track of the usage.

31. 31  1998 Morgan Kaufmann Publishers Performance 0% 3% 6% 9% 12% 15% Eight-wayFour-wayTwo-wayOne-way 1 KB 2 KB 4 KB 8 KB M i s s r a t e Associativity 16 KB 32 KB 64 KB 128 KB

32. 32  1998 Morgan Kaufmann Publishers Decreasing miss penalty with multilevel caches Add a second level cache: –often primary cache is on the same chip as the processor –use SRAMs to add another cache above primary memory (DRAM) –miss penalty goes down if data is in 2nd level cache Example: –CPI of 1.0 on a 500Mhz machine with a 5% miss rate, 200ns DRAM access –Adding 2nd level cache with 20ns access time decreases miss rate to 2% Using multilevel caches: –try and optimize the hit time on the 1st level cache –try and optimize the miss rate on the 2nd level cache

33. 33  1998 Morgan Kaufmann Publishers Virtual Memory: Motivations To allow efficient and safe sharing of memory among multiple programs. To remove the programming burdens of a small, limited amount of main memory.

34. 34  1998 Morgan Kaufmann Publishers Virtual Memory Main memory can act as a cache for the secondary storage (disk) Advantages: –illusion of having more physical memory –program relocation –protection

35. 35  1998 Morgan Kaufmann Publishers Pages: virtual memory blocks Page faults: the data is not in memory, retrieve it from disk –huge miss penalty, thus pages should be fairly large (e.g., 4KB) –reducing page faults is important (LRU is worth the price) –can handle the faults in software instead of hardware –using write-through is too expensive so we use write-back

36. 36  1998 Morgan Kaufmann Publishers Placing a Page and Finding It Again We want the ability to use a clever and flexible replacement scheme. We want to reduce page fault rate. Fully-associative placement serves our purposes. But full search is impractical, so we locate pages by using a full table that indexes the memory. ==> page table (resides in memory) Each program has it own page table, which maps the virtual address space of that program to main memory.

37. 37  1998 Morgan Kaufmann Publishers Page Table Register le register Page table 20 12 18 31 30 29 28 27 15 14 13 12 11 10 9 8 3 2 1 0 29 28 2715 14 13 12 11 10 9 8 3 2 1 0

38. 38  1998 Morgan Kaufmann Publishers Process The page table, together with the program counter and the registers, specifies the state of a program. If we want to allow another program to use the CPU, we must save this state. We often refer to this state as a process. A process is considered active when it’s in possession of the CPU.

39. 39  1998 Morgan Kaufmann Publishers Dealing With Page Faults When the valid bit for a virtual page is off, a page fault occurs. The operating system takes over, and the transfer is done with the exception mechanism. The OS must find the page in the next level of hierarchy, and decide where to place the requested page in the main memory. LRU policy is often used.

40. 40  1998 Morgan Kaufmann Publishers Page Tables

41. 41  1998 Morgan Kaufmann Publishers Page Tables

42. 42  1998 Morgan Kaufmann Publishers Making Address Translation Fast A cache for address translations: translation-lookaside buffer (TLB) age or disk address Physical memory Disk storage

43. 43  1998 Morgan Kaufmann Publishers Integrating VM, TLBs and Caches rty Tag TLB hit Physical page number Physical address tag TLB Physical address 31 30 29 15 14 13 12 11 10 9 8 3 2 1 0

44. 44  1998 Morgan Kaufmann Publishers TLBs and caches

45. 45  1998 Morgan Kaufmann Publishers Implementing Protection with Virtual Memory The OS takes care of this. Hardware need to provide at least three capabilities: –support at least two modes that indicate whether the running process is a user process or an OS process (kernel process, supervisor process, executive process) –provide a portion of the CPU state that a user process can read but not write. –Provide mechanisms whereby the CPU can go from the user mode to supervisor mode.

46. 46  1998 Morgan Kaufmann Publishers A Common Framework for Memory Hierarchies Question 1: Where can a block be placed? Question 2: How is a block found? Question 3: Which block should be replaced on a cache miss? Question 4: What happens on a Write?

47. 47  1998 Morgan Kaufmann Publishers The Three Cs Compulsory misses Capacity misses Conflict misses

48. 48  1998 Morgan Kaufmann Publishers Modern Systems Very complicated memory systems:

49. 49  1998 Morgan Kaufmann Publishers Processor speeds continue to increase very fast — much faster than either DRAM or disk access times Design challenge: dealing with this growing disparity Trends: –synchronous SRAMs (provide a burst of data) –redesign DRAM chips to provide higher bandwidth or processing –restructure code to increase locality –use prefetching (make cache visible to ISA) Some Issues