1. Exploiting Locality in DRAM Xiaodong Zhang College of William and Mary

2. Where is Locality in DRAM? l DRAM is the center of memory hierarchy: l High density and high capacity l Low cost but slow access (compared to SRAM) § A cache miss has been considered as a constant delay for long time. This is wrong. l Non-uniform access latencies exist within DRAM l Row-buffer serves as a fast cache in DRAM l Its access patterns here has been paid little attention. l Reusing buffer data minimizes the DRAM latency. § Larger buffers in DRAM for more locality.

3. Outline l Exploiting locality in Row Buffers l Analysis of access patterns. l A solution to eliminate conflict misses. § Cached DRAM (CDRAM) l Design and its performance evaluation. l Large off-chip cache design by CDAM l Major problems of L3 caches. l Address the problems by CDRAM. § Memory access scheduling l A case for fine grain scheduling.

4. CPU Registers L1 TLB L3 L2 Row buffer DRAM Bus adapter Controller buffer Buffer cache CPU-memory bus I/O bus I/O controller disk Disk cache TLB registers L1 L2 L3 Controller buffer Buffer cache disk cache Row buffer Locality Exploitation in Row Buffer

5. Exploiting the Locality in Row Buffers l Zhang, et. al., Micro-33, 2000, (W&M) l Contributions of this work: l looked into the access patterns in row buffers. l found the reason behind misses in the row buffer. l proposed an effective solution to minimize the misses. § The interleaving technique in this paper was adopted by Sun UltralSPARC IIIi Processor series.

6. DRAM Access = Latency + Bandwidth Time Precharge Row Access Bus bandwidth time DRAM Core Row Buffer Processor Column Access DRAMLatency Row buffer misses come from a sequence of accesses to different pages in the same bank.

7. Nonuniform DRAM Access Latency l Case 1: Row buffer hit (20+ ns) l Case 2: Row buffer miss (core is precharged, 40+ ns ) l Case 3: Row buffer miss (not precharged, ≈ 70 ns) prechargerow accesscol. access row accesscol. access

8. Amdahl’s Law applies in DRAM  As the bandwidth improves, DRAM latency will decide cache miss penalty.  Time (ns) to fetch a 128-byte cache block: latency bandwidth

9. Row Buffer Locality Benefit Objective: serve memory requests without accessing the DRAM core as much as possible. Reduce latency by up to 67%.

10. Row Buffer Misses are Surprisingly High §Standard configuration l Conventional cache mapping l Page interleaving for DRAM memories l 32 DRAM banks, 2KB page size l SPEC95 and SPEC2000 §What is the reason behind this?

11. Conventional Page Interleaving Page 0Page 1Page 2Page 3 Page 4Page 5Page 6Page 7 ………… Bank 0 Address format Bank 1Bank 2Bank 3 page indexpage offsetbank rpk

12. Conflict Sharing in Cache/DRAM cache tagcache set indexblock offset page indexpage offset tsb bank rpk cache-conflicting: same cache index, different tags. row-buffer conflicting: same bank index, different pages. address mapping: bank index  cache set index Property:  x  y, x and y conflict on cache  also on row buffer. page: cache:

13. Sources of Misses l Symmetry: invariance in results under transformations. l Address mapping symmetry propogates conflicts from cache address to memory address space: Cache-conflicting addresses/misses are also row-buffer conflicting addresses/misses. Cache write-back address conflicts with the missed block. Upon a miss, if the replaced cache block is dirty, it must be written back to memory before the missed block is loaded. The conflict between the dirty block address and the missed block address cause a row-buffer miss. As a sequence of replacement on dirty cache blocks happens, so do the write-back conflicts in row-buffer.

14. Breaking the Symmetry by Permutation-based Page Interleaving k XOR k page index page offset new bank k page offset indexbank L2 Cache tag

15. Permutation Property (1) §Conflicting addresses are distributed onto different banks memory banks 0000 0001 0010 0011 0100 0101 0110 0111 1010 1011 Permutation-based interleaving 10111010 1001 1000 1010 L2 Conflicting addresses xor Different bank indexes Conventional interleaving Same bank index

16. Permutation Property (2) §The spatial locality of memory references is preserved. memory banks 0000 0001 0010 0011 0100 0101 0110 0111 1010 1011 10001010 10001010 10001010 10001010 …… Within one page Permutation-based interleaving Conventional interleaving Same bank index xor Same bank index

17. Permutation Property (3) §Pages are uniformly mapped onto ALL memory banks. §P: page, C: the number of pages the (L2/L3) cache holds. C+1P 2C+2P bank 0bank 1bank 2bank 3 C 2C+3P C+3P 2C 01P2P3P C+2P 2C+1P 4P5P6P7P ………… C+5PC+4PC+7PC+6P ………… 2C+6P2C+7P2C+4P2C+5P …………

18. A Solution of ``Swap” l DEC architects ``swap” partial bits of L2 tag and partial bits of the page offset for the AlphaStation 600 5-series. (Digital Technical Journal, 1995). § An optimal number of swapped bits was tested by Wong and Baer (Washington, 97) § We showed why this only slightly solves the problem.

19. Row-buffer Miss Rates

20. Comparison of Memory Stall Times

21. Measuring IPC (#instructions per cycle)

22. Where to Break the Symmetry? l Break the symmetry at the bottom level (DRAM address) is most effective: l Far away from the critical path (little overhead) l Reduce the both address conflicts and write-back conflicts. l Our experiments confirm this (30% difference).

23. Impact to Commercial Systems l Critically show the address mapping problem in Compaq XP1000 series with an effective solution. l Our method has been adopted in Sun Ultra SPARC IIIi processor series, called XOR interleaving. l Chief architect Kevin Normoyle had intensive discussions with us for the adoption in 2001. l The results in the Micro-33 paper on ``conflict propagation”, and ``write-back conflicts” are quoted in the Sun Ultra SPARC Product Manuals. l Sun Microsystems has formally acknowledged our research contribution to their products.

24. Outline

25. Can We Exploit More Locality in DRAM? l Cached DRAM: adding a small on-memory cache in the memory core. l Exploiting the locality in main memory by the cache. l High bandwidth between the cache and memory core. l Fast response to single memory request hit in the cache. l Pipelining multiple memory requests starting from the memory controller via the memory bus, the cache, and the DRAM core (if on-memory cache misses happen).

26. Cached DRAM CPU L1 Cache L2 Cache Memory Bus On Memory Cache DRAM Core Cached DRAM Low bandwidth in cache line per bus cycle High bandwidth in page per internal bus cycle

27. Improvement of IPC (# of instructions per cycle)

28. Cached DRAM vs. XOR Interleaving (16 * 4 KB on-memory cache for CDRAM, 32 * 2 KB row buffers for XOR interleaving among 32 banks)

29. Cons and Pros of CDRAM over xor Interleaving l Merits: l High hits in on-memory cache due to high associativity. l The cache can be accessed simultaneously with DRAM. l More cache blocks than the number of memory banks. l Limits: l Requires an additional chip area in DRAM core and additional management circuits.

30. Outline

31. Large Off-chip Caches by CDRAM l Large and off-chip L3 caches are commonly used to reduce memory latency. l It has some limits for large memory intensive applications: l The size is still limited (less than 10 MB). l Access latency is large (10+ times over on-chip cache) l Large volume of L3 tags (tag checking time ∞ log (tag size) l Tags are stored off-chip. l Study shows that L3 can degrade performance for some applications (DEC Report 1996).

32. Can CDRAM Address L3 Problems? l What happens if L3 is replaced by CDRAM? l The size of CDRAM is sufficiently large, however, l How could its average latency be comparable or even lower than L3 cache? l The challenge is to reduce the access latency to this huge ``off-chip cache”. l ``Cached DRAM Cache” (CDC) addresses the L3 problem, by Zhang et. al. published in IEEE Transactions on Computers in 2004. (W&M)

33. Cached DRAM Cache as L3 in Memory Hierarchy L1 Inst CacheL1 Data Cache L2 Unified Cache CDC tag cache and predictor CDC-DRAM CDC-cache Memory bus DRAM main memory

34. How is the Access Latency Reduced? l The tags of the CDC cache are stored on-chip. l Demanding a very small storage. l High hits in CDC cache due to high locality of L2 miss streams. l Unlike L3, the CDC is not between L2 and DRAM. l It is in parallel with the DRAM memory. l An L2 miss can either go to CDC or DRAM via different buses. l Data fetching in CDC and DRAM can be done independently. l A predictor is built on-chip using a global history register. l Determine if a L2 miss will be a hit/miss in CDC. l The accuracy is quite high.

35. Advantages and Performance Gains l Unique advantages l Large capacity, equivalent to the DRAM size, and l Low average latency by (1) exploiting locality in CDC-cache, (2) fast on-chip tag checking for CDC-cache data, (3) accurate prediction of hit/miss in CDC. l Performance of SPEC2000 l Outperforms L3 organization by up to 51%. l Unlike L3, CDC does not degrade performance of any. l The average performance improvement is 25%.

36. Performance Evaluation by SPEC2000fp

37. Outline

38. Memory Access Scheduling l Objectives: l Fully utilize the memory resources, such as buses and concurrency of operations in banks and transfers. l Minimizing the access time by eliminating potential access contention. l Access orders based on priorities make a significant performance difference. l Improving functionalities in Memory Controller.

39. CPU FIFO Stream Buffer Unit Address Mapping Unit Cache Main Memory Memory Scheduling Unit Memory Controller Memory accesses issued in the requested order Memory accesses issued in an “optimal” order

40. Basic Functions of Memory Controller l Where is it? l A hardware logic directly connected to CPU, which generates necessary signals to control the read/write, and address mapping in the memory, and interface other with other system components (CPU, cache). l What does it do specifically? l Pipelining and buffering the requests l Memory address mapping (e.g. XOR interleaving) l Reorder the memory accesses to improve performance.

41. Complex Configuration of Memory Systems l Multi-channel memory systems (e.g. Rambus) l Each channel connect multiple memory devices. l Each device consists multiple memory banks. l Current operations among channels and banks. l How to utilize rich multi-channel resources? l Maximizing the concurrent operations. l Deliver a cache line with critical sub-block first.

42. Device 0 Device D-1 …… Bank 0 Bank B-1 …… …… Channel C-1 …… Channel 0 Multi-channel Memory Systems CPU /L1 L2

43. Partitioning A Cache Line into sub-blocks §Smaller sub-block size  shorter latency for critical sub-blocks §DRAM system: minimal request length  Sub-block size = smallest granularity available for Direct Rambus system a cache miss request multiple DRAM Requests (in the same bank)

44. Mapping Sub-blocks onto Multi-channels Evenly distribute sub-blocks to all channels  aggregate bandwidth for each cache request channel 0channel 1 a cache line fill request

45. Priority Ranks of Sub-blocks §Read-bypass-write: a ``read” is in the critical path and requires less delay than write. A memory ``write” can be overlapped with others operations. §Hit-first: row buffer hit. Get it before it is replaced. §Ranks for read/write  Critical: critical load sub-requests of cache read misses  Load: non-critical load sub-requests of cache read misses  Store: load sub-requests for cache write misses §In-order: other serial accesses.

46. Existing Scheduling Methods for MC l Gang scheduling: (Lin, et. al., HPCA’01, Michigan) l Upon a cache miss, all the channels are used to deliver. l Maximize concurrent operations among multi-channels. l Effective to a single miss, but not for multiple misses (cache lines have to be delivered one by one). l No consideration for sub-block priority. l Burst scheduling (Cuppu, et. al., ISCA’01, Maryland) l One cache line per channel, and reorder the sub-blocks in each. l Effective to multiple misses, not to a single or small number of misses (under utilizing concurrent operations in multi-channels).

47. Fine Grain Memory Access Scheduling l Zhu, et., al., HPCA’02 (W&M). l Sub-block and its priority based scheduling. l All the channels are used at a time. l Always deliver the high priority blocks first. l Priority of each critical sub-block is a key.

48. Fine Grain Both P&C. Burst Use priority,but not all channels. Gang Use all channels But no priority. Advantages of Fine Grain Scheduling A5 A0 A1 A2 A3 A4 A6 A7 B0 B1 B2 B3 B4 B5 B6 B7 A0A1A2A3A4A5A6A7 B0B1B2B3B4B5B6B7 A0 A1 A2 A3 A4 A5 A6 A7 B0 B1 B2 B3 B4 B5 B6 B7 A0 A1 A2 A3 A4 A5 A6 A7 B0 B1 B2 B3 B4 B5 B6 B7

49. Experimental Environment §Simulator l SimpleScalar 3.0b l An event-driven simulation of a multi- channel Direct Rambus DRAM system §Benchmark l SPEC CPU2000 §Key parameters l Processor: 2GHz, 4-issue l MSHR: 16 entries l L1 cache : 4-way 64KB I/D l L2 cache: 4-way 1MB, 128B block l Channel: 2 or 4 l Device: 4 / channel l Bank: 32 / device l Length of packets: 16 B l Precharge: 20 ns l Row access: 20 ns l Column access: 20 ns

50. Burst Phase in Miss Streams

51. Clustering of Multiple Accesses

52. Percentages of Critical Sub-blocks

53. Waiting Time Distribution

54. Critical Sub-block Distribution in Channels

55. Performance Improvement: Fine Grain Over Gang Scheduling

56. Performance Improvement: Fine Grain Over Burst Scheduling

57. 2-channel Fine Grain Vs. 4-channel Gang & Burst Scheduling

58. Summary of Memory Access Scheduling §Fine-grain priority scheduling l Granularity: sub-block based. l Mapping schemes: utilize all the channels. l Scheduling policies: priority based. §Outperforms Gang & Burst Scheduling l Effective utilizing available bandwidth and concurrency l Reducing average waiting time for cache miss requests l Reducing processor stall time for memory accesses

59. Conclusion l High locality exists in cache miss streams. l Exploiting locality in row buffers can make a great performance difference. l Cached DRAM can further exploit the locality in DRAM. l CDCs can serve as large and low overhead off-chip caches. l Memory access scheduling plays a critical role. l Exploiting locality in DRAM is very unique. l Direct and positive impact to commercial product. l The locality in DRAM has been ignored for long time. l Impact to architecture and computer organization teaching.