1. Parallelism Lecture notes from MKP and S. Yalamanchili

2. (2) Introduction Goal: Higher performance through parallelism Job-level (process-level) parallelism  High throughput for independent jobs Application-level parallelism  Single program run on multiple processors Multicore microprocessors  Chips with multiple processors (cores)  Support for both job level and application-level parallelism

3. (3) Core Count Roadmap: AMD From anandtech.com

4. (4) Core Count: NVIDIA 1536 cores at 1GHz All cores are not created equal Need to understand the programming model

5. (5) Hardware and Software Hardware  Serial: e.g., Pentium 4  Parallel: e.g., quad-core Xeon e5345 Software  Sequential: e.g., matrix multiplication  Concurrent: e.g., operating system Sequential/concurrent software can run on serial/parallel hardware  Challenge: making effective use of parallel hardware

6. (6) Parallel Programming Parallel software is the problem Need to get significant performance improvement  Otherwise, just use a faster uniprocessor, since it ’ s easier! Difficulties  Partitioning  Coordination  Communications overhead

7. (7) Amdahl ’ s Law Sequential part can limit speedup Example: 100 processors, 90× speedup?  T new = T parallelizable /100 + T sequential   Solving: F parallelizable = 0.999 Need sequential part to be 0.1% of original time

8. (8) Scaling Example Workload: sum of 10 scalars, and 10 × 10 matrix sum  Speed up from 10 to 100 processors Single processor: Time = (10 + 100) × t add 10 processors  Time = 10 × t add + 100/10 × t add = 20 × t add  Speedup = 110/20 = 5.5 (55% of potential) 100 processors  Time = 10 × t add + 100/100 × t add = 11 × t add  Speedup = 110/11 = 10 (10% of potential) Idealized model  Assumes load can be balanced across processors

9. (9) Scaling Example (cont) What if matrix size is 100 × 100? Single processor: Time = (10 + 10000) × t add 10 processors  Time = 10 × t add + 10000/10 × t add = 1010 × t add  Speedup = 10010/1010 = 9.9 (99% of potential) 100 processors  Time = 10 × t add + 10000/100 × t add = 110 × t add  Speedup = 10010/110 = 91 (91% of potential) Idealized model  Assuming load balanced

10. (10) Strong vs Weak Scaling Strong scaling: problem size fixed  As in example Weak scaling: problem size proportional to number of processors  10 processors, 10 × 10 matrix oTime = 20 × t add  100 processors, 32 × 32 matrix oTime = 10 × t add + 1000/100 × t add = 20 × t add  Constant performance in this example  For a fixed size system grow the number of processors to improve performance

11. (11) What We Have Seen §3.6: Parallelism and Computer Arithmetic  Associativity and bit level parallelism §4.10: Parallelism and Advanced Instruction- Level Parallelism  Recall multi-instruction issue §6.9: Parallelism and I/O:  Redundant Arrays of Inexpensive Disks Now we will look at categories in computation  classification

12. (12) Concurrency and Parallelism Concurrent access to shared data must be controlled for correctness Programming models? Image from futurelooks.com Each core can operate concurrently and in parallel Multiple threads may operate in a time sliced fashion on a single core

13. (13) Instruction Level Parallelism (ILP) IFIDMEMWB Single (program) thread of execution Issue multiple instructions from the same instruction stream Average CPI<1 Often called out of order (OOO) cores Multiple instructions in EX at the same time

14. (14) The P4 Microarchitecture From, “The Microarchitecture of the Pentium 4 Processor 1,” G. Hinton et.al, Intel Technology Journal Q1, 2001

15. (15) ILP Wall - Past the Knee of the Curve? “Effort” Performance Scalar In-Order Moderate-Pipe Superscalar/OOO Very-Deep-Pipe Aggressive Superscalar/OOO Made sense to go Superscalar/OOO: good ROI Very little gain for substantial effort Source: G. Loh

16. (16) Thread Level Parallelism (TLP) Multiple threads of execution Exploit ILP in each thread Exploit concurrent execution across threads

17. (17) Instruction and Data Streams Taxonomy due to M. Flynn Data Streams SingleMultiple Instruction Streams SingleSISD: Intel Pentium 4 SIMD: SSE instructions of x86 MultipleMISD: No examples today MIMD: Intel Xeon e5345 SPMD: Single Program Multiple Data A parallel program on a MIMD computer where each instruction stream is identical Conditional code for different processors

18. (18) Programming Model: Multithreading Performing multiple threads of execution in parallel  Replicate registers, PC, etc.  Fast switching between threads Fine-grain multithreading  Switch threads after each cycle  Interleave instruction execution  If one thread stalls, others are executed Coarse-grain multithreading  Only switch on long stall (e.g., L2-cache miss)  Simplifies hardware, but doesn ’ t hide short stalls (eg, data hazards)

19. (19) Conventional Multithreading Zero-overhead context switch Duplicated contexts for threads 0:r0 0:r7 1:r0 1:r7 2:r0 2:r7 3:r0 3:r7 CtxtPtr Memory (shared by threads) Register file

20. (20) Simultaneous Multithreading In multiple-issue dynamically scheduled processor  Schedule instructions from multiple threads  Instructions from independent threads execute when function units are available  Within threads, dependencies handled by scheduling and register renaming Example: Intel Pentium-4 HT  Two threads: duplicated registers, shared function units and caches  Known as Hyperthreading in Intel terminology

21. (21) 21 Hyper-threading Implementation of Hyper-threading adds less that 5% to the chip area Principle: share major logic components by adding or partitioning buffering logic Processor Execution Resources Arch State Processor Execution Resources Arch State 2 CPU Without Hyper-threading2 CPU With Hyper-threading

22. (22) Multithreading Example

23. (23) Shared Memory SMP: shared memory multiprocessor  Hardware provides single physical address space for all processors  Synchronize shared variables using locks  Memory access time oUMA (uniform) vs. NUMA (nonuniform)

24. (24) Example: Communicating Threads Producer The Producer calls while (1) { while (count == BUFFER_SIZE) ; // do nothing // add an item to the buffer ++count; buffer[in] = item; in = (in + 1) % BUFFER_SIZE; } Producer Consumer

25. (25) Example: Communicating Threads Consumer The Consumer calls while (1) { while (count == 0) ; // do nothing // remove an item from the buffer --count; item = buffer[out]; out = (out + 1) % BUFFER_SIZE; } Producer Consumer

26. (26) Uniprocessor Implementation count++ could be implemented as register1 = count; register1 = register1 + 1; count = register1; count-- could be implemented as register2 = count; register2 = register2 – 1; count = register2; Consider this execution interleaving: S0: producer execute register1 = count {register1 = 5} S1: producer execute register1 = register1 + 1 {register1 = 6} S2: consumer execute register2 = count {register2 = 5} S3: consumer execute register2 = register2 - 1 {register2 = 4} S4: producer execute count = register1 {count = 6 } S5: consumer execute count = register2 {count = 4}

27. (27) Synchronization We need to prevent certain instruction interleavings  Or at least be able to detect violations! Some sequence of operations (instructions) must happen atomically  E.g., register1 = count; register1 = register1 + 1; count = register1;  atomic operations/instructions

28. (28) Synchronization Two processors sharing an area of memory  P1 writes, then P2 reads  Data race if P1 and P2 don’t synchronize oResult depends of order of accesses Hardware support required  Atomic read/write memory operation  No other access to the location allowed between the read and write Could be a single instruction  E.g., atomic swap of register ↔ memory  Or an atomic pair of instructions

29. (29) Synchronization in MIPS Load linked: ll rt, offset(rs) Store conditional: sc rt, offset(rs)  Succeeds if location not changed since the ll oReturns 1 in rt  Fails if location is changed oReturns 0 in rt Example: atomic swap (to test/set lock variable) try: add $t0,$zero,$s4 ;copy exchange value ll $t1,0($s1) ;load linked sc $t0,0($s1) ;store conditional beq $t0,$zero,try ;branch store fails add $s4,$zero,$t1 ;put load value in $s4

30. (30) Cache Coherence A shared variable may exist in multiple caches Multiple copies to improve latency This is a really a synchronization problem

31. (31) Cache Coherence Problem Suppose two CPU cores share a physical address space  Write-through caches Time step EventCPU A’s cache CPU B’s cache Memory 00 1CPU A reads X00 2CPU B reads X000 3CPU A writes 1 to X101

32. (32) Example (Writeback Cache) P Cache Memory P X= -100 Cache P X= -100 X= 505 Rd ? X= -100 Rd ? Courtesy H. H. Lee

33. (33) Coherence Defined Informally: Reads return most recently written value Formally:  P writes X; P reads X (no intervening writes)  read returns written value  P 1 writes X; P 2 reads X (sufficiently later)  read returns written value oc.f. CPU B reading X after step 3 in example  P 1 writes X, P 2 writes X  all processors see writes in the same order oEnd up with the same final value for X

34. (34) Cache Coherence Protocols Operations performed by caches in multiprocessors to ensure coherence  Migration of data to local caches oReduces bandwidth for shared memory  Replication of read-shared data oReduces contention for access Snooping protocols  Each cache monitors bus reads/writes Directory-based protocols  Caches and memory record sharing status of blocks in a directory

35. (35) Invalidating Snooping Protocols Cache gets exclusive access to a block when it is to be written  Broadcasts an invalidate message on the bus  Subsequent read in another cache misses oOwning cache supplies updated value CPU activityBus activityCPU A’s cache CPU B’s cache Memory 0 CPU A reads XCache miss for X00 CPU B reads XCache miss for X000 CPU A writes 1 to XInvalidate for X10 CPU B read XCache miss for X111

36. (36) Programming Model: Message Passing Each processor has private physical address space Hardware sends/receives messages between processors

37. (37) Parallelism Write message passing programs Explicit send and receive of data  Rather than accessing data in shared memory send() receive() send() receive() Process 2

38. (38) Loosely Coupled Clusters Network of independent computers  Each has private memory and OS  Connected using I/O system oE.g., Ethernet/switch, Internet Suitable for applications with independent tasks  Web servers, databases, simulations, … High availability, scalable, affordable Problems  Administration cost (prefer virtual machines)  Low interconnect bandwidth oc.f. processor/memory bandwidth on an SMP

39. (39) High Performance Computing zdnet.com The dominant programming model is message passing Scales well but requires programmer effort Science problems have fit this model well to date theregister.co.uk

40. (40) Grid Computing Separate computers interconnected by long- haul networks  E.g., Internet connections  Work units farmed out, results sent back Can make use of idle time on PCs  E.g., SETI@home, World Community Grid

41. (41) Programming Model: SIMD Operate elementwise on vectors of data  E.g., MMX and SSE instructions in x86 oMultiple data elements in 128-bit wide registers All processors execute the same instruction at the same time  Each with different data address, etc. Simplifies synchronization Reduced instruction control hardware Works best for highly data-parallel applications Data Level Parallelism

42. (42) SIMD Co-Processor Graphics and media processing operates on vectors of 8-bit and 16-bit data  Use 64-bit adder, with partitioned carry chain oOperate on 8×8-bit, 4×16-bit, or 2×32-bit vectors  SIMD (single-instruction, multiple-data) 4x16-bit2x32-bit

43. (43) History of GPUs Early video cards  Frame buffer memory with address generation for video output 3D graphics processing  Originally high-end computers (e.g., SGI)  Moore ’ s Law  lower cost, higher density  3D graphics cards for PCs and game consoles Graphics Processing Units  Processors oriented to 3D graphics tasks  Vertex/pixel processing, shading, texture mapping, rasterization

44. (44) Graphics in the System

45. (45) GPU Architectures Processing is highly data-parallel  GPUs are highly multithreaded  Use thread switching to hide memory latency oLess reliance on multi-level caches  Graphics memory is wide and high-bandwidth Trend toward general purpose GPUs  Heterogeneous CPU/GPU systems  CPU for sequential code, GPU for parallel code Programming languages/APIs  DirectX, OpenGL  C for Graphics (Cg), High Level Shader Language (HLSL)  Compute Unified Device Architecture (CUDA )

46. (46) Example: NVIDIA Tesla Streaming multiprocessor 8 × Streaming processors

47. (47) Compute Unified Device Architecture http://developer.nvidia.com/cuda-education-training For access to CUDA tutorials Bulk synchronous processing (BSP) execution model 47

48. (48) Example: NVIDIA Tesla Streaming Processors  Single-precision FP and integer units  Each SP is fine-grained multithreaded Warp: group of 32 threads  Executed in parallel, SIMD style o8 SPs × 4 clock cycles  Hardware contexts for 24 warps oRegisters, PCs, …

49. (49) Classifying GPUs Does not fit nicely into SIMD/MIMD model  Conditional execution in a thread allows an illusion of MIMD oBut with performance degradation oNeed to write general purpose code with care Static: Discovered at Compile Time Dynamic: Discovered at Runtime Instruction-Level Parallelism VLIWSuperscalar Data-Level Parallelism SIMD or VectorTesla Multiprocessor Really Single Instruction Multiple Thread (SIMT)

50. (50) Vector Processors Highly pipelined function units Stream data from/to vector registers to units  Data collected from memory into registers  Results stored from registers to memory Example: Vector extension to MIPS  32 × 64-element registers (64-bit elements)  Vector instructions olv, sv : load/store vector oaddv.d : add vectors of double oaddvs.d : add scalar to each element of vector of double Significantly reduces instruction-fetch bandwidth

51. (51) Cray-1: Vector Machine Mid 70’s – principles have not changed Aggregate operations defined on vectors Vector ISAs operating on vector instruction sets Load/store ISA ala MIPS Often confused with SIMD From pg-server.csc.ncsu.edu

52. (52) Vector vs. Scalar Vector architectures and compilers  Simplify data-parallel programming  Explicit statement of absence of loop-carried dependences oReduced checking in hardware  Regular access patterns benefit from interleaved and burst memory  Avoid control hazards by avoiding loops More general than ad-hoc media extensions (such as MMX, SSE)  Better match with compiler technology

53. (53) Interconnection Networks Network topologies  Arrangements of processors, switches, and links BusRing 2D Mesh N-cube (N = 3) Fully connected

54. (54) Network Characteristics Performance  Latency per message (unloaded network)  Throughput oLink bandwidth oTotal network bandwidth oBisection bandwidth  Congestion delays (depending on traffic) Cost Power Routability in silicon

55. (55) Modeling Performance Assume performance metric of interest is achievable GFLOPs/sec  Measured using computational kernels from Berkeley Design Patterns Arithmetic intensity of a kernel  FLOPs per byte of memory accessed For a given computer, determine  Peak GFLOPS (from data sheet)  Peak memory bytes/sec (using Stream benchmark)

56. (56) Roofline Diagram Attainable GPLOPs/sec = Max ( Peak Memory BW × Arithmetic Intensity, Peak FP Performance )

57. (57) Comparing Systems Example: Opteron X2 vs. Opteron X4  2-core vs. 4-core, 2× FP performance/core, 2.2GHz vs. 2.3GHz  Same memory system To get higher performance on X4 than X2 Need high arithmetic intensity Or working set must fit in X4’s 2MB L-3 cache

58. (58) Optimizing Performance Optimize FP performance  Balance adds & multiplies  Improve superscalar ILP and use of SIMD instructions Optimize memory usage  Software prefetch oAvoid load stalls  Memory affinity oAvoid non-local data accesses

59. (59) Optimizing Performance Choice of optimization depends on arithmetic intensity of code Arithmetic intensity is not always fixed May scale with problem size Caching reduces memory accesses Increases arithmetic intensity

60. (60) Four Example Systems 2 × quad-core Intel Xeon e5345 (Clovertown) 2 × quad-core AMD Opteron X4 2356 (Barcelona)

61. (61) Four Example Systems 2 × oct-core IBM Cell QS20 2 × oct-core Sun UltraSPARC T2 5140 (Niagara 2)

62. (62) And Their Rooflines Kernels  SpMV (left)  LBHMD (right) Some optimizations change arithmetic intensity x86 systems have higher peak GFLOPs  But harder to achieve, given memory bandwidth

63. (63) Pitfalls Not developing the software to take account of a multiprocessor architecture  Example: using a single lock for a shared composite resource oSerializes accesses, even if they could be done in parallel oUse finer-granularity locking

64. (64) Concluding Remarks Goal: higher performance by using multiple processors Difficulties  Developing parallel software  Devising appropriate architectures Many reasons for optimism  Changing software and application environment  Chip-level multiprocessors with lower latency, higher bandwidth interconnect An ongoing challenge for computer architects!

65. (65) Study Guide Be able to explain the following concepts ILP, MT, SMT, TLP, DLP, MIMD, SIMD, SISD Explain the roofline model of performance Use of Amdahl’s Law in demonstrating the limits of scaling What is the impact of a sequence of read/write operations on shared data?  Cache coherence How does ILP differ from SMT How does SIMD differ from vector? What is the difference between weak vs. strong scaling?