1. Unified Parallel C (UPC) Costin Iancu The Berkeley UPC Group: C. Bell, D. Bonachea, W. Chen, J. Duell, P. Hargrove, P. Husbands, C. Iancu, R. Nishtala, M. Welcome, K. Yelick http://upc.lbl.gov Slides edited by K. Yelick, T. El-Ghazawi, P. Husbands, C.Iancu

2. Context Most parallel programs are written using either: –Message passing with a SPMD model (MPI) Usually for scientific applications with C++/Fortran Scales easily –Shared memory with threads in OpenMP, Threads+C/C++/F or Java Usually for non-scientific applications Easier to program, but less scalable performance Partitioned Global Address Space (PGAS) Languages take the best of both –SPMD parallelism like MPI (performance) –Local/global distinction, i.e., layout matters (performance) –Global address space like threads (programmability) 3 Current languages: UPC (C), CAF (Fortran), and Titanium (Java) 3 New languages: Chapel, Fortress, X10

3. Partitioned Global Address Space Shared memory is logically partitioned by processors Remote memory may stay remote: no automatic caching implied One-sided communication: reads/writes of shared variables Both individual and bulk memory copies Some models have a separate private memory area Distributed array generality and how they are constructed Shared Global address space X[0] Private ptr: X[1]X[P] Thread 0 Thread 1 Thread n

4. Partitioned Global Address Space Languages Explicitly-parallel programming model with SPMD parallelism –Fixed at program start-up, typically 1 thread per processor Global address space model of memory –Allows programmer to directly represent distributed data structures Address space is logically partitioned –Local vs. remote memory (two-level hierarchy) Programmer control over performance critical decisions –Data layout and communication Performance transparency and tunability are goals –Initial implementation can use fine-grained shared memory

5. Current Implementations A successful language/library must run everywhere UPC –Commercial compilers: Cray, SGI, HP, IBM –Open source compilers: LBNL/UCB (source-to-source), Intrepid (gcc) CAF –Commercial compilers: Cray –Open source compilers: Rice (source-to-source) Titanium –Open source compilers: UCB (source-to-source) Common tools –Open64 open source research compiler infrastructure –ARMCI, GASNet for distributed memory implementations –Pthreads, System V shared memory

6. Talk Overview UPC Language Design –Data Distribution (layout, memory management) –Work Distribution (data parallelism) –Communication (implicit,explicit, collective operations) –Synchronization (memory model, locks) Programming in UPC –Performance (one-sided communication) –Application examples: FFT, PC –Productivity (compiler support) –Performance tuning and modeling

7. UPC Overview and Design Unified Parallel C (UPC) is: –An explicit parallel extension of ANSI C with common and familiar syntax and semantics for parallel C and simple extensions to ANSI C –A partitioned global address space language (PGAS) –Based on ideas in Split-C, AC, and PCP Similar to the C language philosophy –Programmers are clever and careful, and may need to get close to hardware to get performance, but can get in trouble SPMD execution model: ( THREADS, MYTHREAD), static vs. dynamic threads

8. Data Distribution

9. Shared Global Address Space X[0] Private ptr: X[1]X[P] Thread 0 Thread 1 Thread n ours: mine: Distinction between memory spaces through extensions of the type system ( shared qualifier) shared int ours; shared int X[THREADS]; shared int *ptr; int mine; Data in shared address space: –Static: scalars (T0), distributed arrays –Dynamic: dynamic memory management ( upc_alloc, upc_global_alloc, upc_all_alloc )

10. Data Layout Data layout controlled through extensions of the type system (layout specifiers): –[0] or [] (indefinite layout, all on 1 thread): shared [] int *p; –Empty (cyclic layout) : shared int array[THREADS*M]; –[*] (blocked layout): shared [*] int array[THREADS*M]; –[b] or [b1][b2]…[bn] = [b1*b2*…bn] (block cyclic) shared [B] int array[THREADS*M]; Element array[i] has affinity with thread (i / B) % THREADS Layout determines pointer arithmetic rules Introspection ( upc_threadof, upc_phaseof, upc_blocksize )

11. UPC Pointers Implementation In UPC pointers to shared objects have three fields: –thread number –local address of block –phase (specifies position in the block) Example: Cray T3E implementation Pointer arithmetic can be expensive in UPC PhaseThreadVirtual Address 03738484963 Virtual AddressThreadPhase

12. UPC Pointers LocalShared Private PP ( p1 )PS ( p3 ) Shared SP ( p2 )SS ( p4 ) Where does the pointer point? Where does the pointer reside? int *p1; /* private pointer to local memory */ shared int *p2; /* private pointer to shared space */ int *shared p3; /* shared pointer to local memory */ shared int *shared p4; /* shared pointer to shared space */

13. UPC Pointers int *p1; /* private pointer to local memory */ shared int *p2; /* private pointer to shared space */ int *shared p3; /* shared pointer to local memory */ shared int *shared p4; /* shared pointer to shared space */ Shared Global address space Private p1: Thread 0 Thread 1 Thread n p2: p1: p2: p1: p2: p3: p4: p3: p4: p3: p4: Pointers to shared often require more storage and are more costly to dereference; they may refer to local or remote memory.

14. Common Uses for UPC Pointer Types int *p1; These pointers are fast (just like C pointers) Use to access local data in part of code performing local work Often cast a pointer-to-shared to one of these to get faster access to shared data that is local shared int *p2; Use to refer to remote data Larger and slower due to test-for-local + possible communication int *shared p3; Not recommended shared int *shared p4; Use to build shared linked structures, e.g., a linked list typedef is the UPC programmer’s best friend

15. UPC Pointers Usage Rules Pointer arithmetic supports blocked and non-blocked array distributions Casting of shared to private pointers is allowed but not vice versa ! When casting a pointer-to-shared to a pointer-to- local, the thread number of the pointer to shared may be lost Casting of shared to local is well defined only if the object pointed to by the pointer to shared has affinity with the thread performing the cast

16. Work Distribution

17. Owner computes rule: loop over all, work on those owned by you UPC adds a special type of loop upc_forall(init; test; step; affinity) statement; Programmer indicates the iterations are independent –Undefined if there are dependencies across threads Affinity expression indicates which iterations to run on each thread. It may have one of two types: –Integer: affinity%THREADS == MYTHREAD –Pointer: upc_threadof(affinity) == MYTHREAD Syntactic sugar for: for(i=MYTHREAD; i<N; i+=THREADS) … for(i=0; i<N; i++) if (MYTHREAD == i%THREADS) … Work Distribution upc_forall()

18. Inter-Processor Communication

19. Data Communication Implicit (assignments) shared int *p; … *p = 7; Explicit (bulk synchronous) point-to-point ( upc_memget, upc_memput, upc_memcpy, upc_memset ) Collective operations http://www.gwu.edu/~upc/docs/ http://www.gwu.edu/~upc/docs/ –Data movement: broadcast, scatter, gather, … –Computational: reduce, prefix, … –Interface has synchronization modes: (??) Avoid over-synchronizing (barrier before/after is simplest semantics, but may be unnecessary) Data being collected may be read/written by any thread simultaneously

20. Data Communication The UPC Language Specification V 1.2 does not contain non-blocking communication primitives Extensions for non-blocking communication available in the BUPC implementation UPC V1.2 does not have higher level communication primitives for point-to-point communication. See BUPC extensions for –scatter, gather –VIS Should non-blocking communication be a first class language citizen?

21. Synchronization

22. Point-to-point synchronization: locks –opaque type : upc_lock_t* –dynamically managed: upc_all_lock_alloc, upc_global_lock_alloc Global synchronization: –Barriers (unaligned) : upc_barrier –Split-phase barriers: upc_notify; this thread is ready for barrier do computation unrelated to barrier upc_wait; wait for others to be ready

23. Memory Consistency in UPC The consistency model defines the order in which one thread may see another threads accesses to memory –If you write a program with un-synchronized accesses, what happens? –Does this work? data = … while (!flag) { }; flag = 1; … = data; // use the data UPC has two types of accesses: –Strict: will always appear in order –Relaxed: May appear out of order to other threads Consistency is associated either with a program scope (file, statement) { #pragma upc strict flag = 1; } or with a type shared strict int flag;

24. Sample UPC Code

25. Matrix Multiplication in UPC Given two integer matrices A(NxP) and B(PxM), we want to compute C =A x B. Entries c ij in C are computed by the formula:

26. Serial C code #define N 4 #define P 4 #define M 4 int a[N][P] = {1,2,3,4,5,6,7,8,9,10,11,12,14,14,15,16}, c[N][M]; int b[P][M] = {0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1}; void main (void) { int i, j, l; for (i = 0 ; i<N ; i++) { for (j=0 ; j<M ;j++) { c[i][j] = 0; for (l = 0 ; l<P ; l++) c[i][j] += a[i][l]*b[l][j]; }

27. Domain Decomposition A (N  P) is decomposed row- wise into blocks of size (N  P) / THREADS as shown below: B(P  M) is decomposed column wise into M/ THREADS blocks as shown below: Thread 0 Thread 1 Thread THREADS-1 0.. (N*P / THREADS) -1 (N*P / THREADS)..(2*N*P / THREADS)-1 ((THREADS-1)  N*P) / THREADS.. (THREADS*N*P / THREADS)-1 Columns 0: (M/THREADS)-1 Columns ((THREAD-1)  M)/THREADS:(M-1) Thread 0 Thread THREADS-1 Note: N and M are assumed to be multiples of THREADS Exploits locality in matrix multiplication N P M P

28. UPC Matrix Multiplication Code #include #define N 4 #define P 4 #define M 4 shared [N*P/THREADS] int a[N][P] = {1,..,16}, c[N][M]; // data distribution: a and c are blocked shared matrices shared [M/THREADS] int b[P][M] = {0,1,0,1,…,0,1}; void main (void) { int i, j, l; // private variables upc_forall(i = 0 ; i<N ; i++; &c[i][0]) { //work distribution for (j=0 ; j<M ;j++) { c[i][j] = 0; for (l= 0 ; l<P ; l++) //implicit communication c[i][j] += a[i][l]*b[l][j]; } }

29. UPC Matrix Multiplication With Block Copy #include shared [N*P /THREADS] int a[N][P], c[N][M]; // a and c are blocked shared matrices shared[M/THREADS] int b[P][M]; int b_local[P][M]; void main (void) { int i, j, l; // private variables //explicit bulk communication upc_memget(b_local, b, P*M*sizeof(int)); //work distribution (c aligned with a??) upc_forall(i = 0 ; i<N ; i++; &c[i][0]) { for (j=0 ; j<M ;j++) { c[i][j] = 0; for (l= 0 ; l<P ; l++) c[i][j] += a[i][l]*b_local[l][j]; }

30. Programming in UPC Don’t ask yourself what can my compiler do for me, ask yourself what can I do for my compiler!

31. Principles of Performance Software To minimize the cost of communication Use the best available communication mechanism on a given machine Hide communication by overlapping (programmer or compiler or runtime) Avoid synchronization using data-driven execution (programmer or runtime) Tune communication using performance models when they work (??); search when they don’t (programmer or compiler/runtime)

32. Best Available Communication Mechanism Performance is determined by overhead, latency and bandwidth Data transfer (one-sided communication) is often faster than (two sided) message passing Semantics limit performance –In-order message delivery –Message and tag matching –Need to acquire information from remote host processor –Synchronization (message receipt) tied to data transfer

33. One-Sided vs Two-Sided: Theory A two-sided messages needs to be matched with a receive to identify memory address to put data –Offloaded to Network Interface in networks like Quadrics –Need to download match tables to interface (from host) A one-sided put/get message can be handled directly by a network interface with RDMA support –Avoid interrupting the CPU or storing data from CPU (preposts) address message id data payload one-sided put message two-sided message network interface memory host CPU

34. GASNet: Portability and High-Performance (down is good) GASNet better for overhead and latency across machines UPC Group; GASNet design by Dan Bonachea

35. (up is good) GASNet excels at mid-range sizes: important for overlap GASNet: Portability and High-Performance Joint work with UPC Group; GASNet design by Dan Bonachea (up is good) GASNet at least as high (comparable) for large messages

36. One-Sided vs. Two-Sided: Practice InfiniBand: GASNet vapi-conduit and OSU MVAPICH 0.9.5 Half power point (N ½ ) differs by one order of magnitude This is not a criticism of the implementation! Yelick,Hargrove, Bonachea (up is good) NERSC Jacquard machine with Opteron processors

37. Overlap

38. Hide Communication by Overlapping A programming model that decouples data transfer and synchronization (init, sync) BUPC has several extensions: (programmer) –explicit handle based –region based –implicit handle based Examples: –3D FFT (programmer) –split-phase optimizations (compiler) –automatic overlap (runtime)

39. Performing a 3D FFT NX x NY x NZ elements spread across P processors Will Use 1-Dimensional Layout in Z dimension –Each processor gets NZ / P “planes” of NX x NY elements per plane 1D Partition NX NY Example: P = 4 NZ p0 p1 p2 p3 NZ/P Bell, Nishtala, Bonachea, Yelick

40. Performing a 3D FFT (part 2) Perform an FFT in all three dimensions With 1D layout, 2 out of the 3 dimensions are local while the last Z dimension is distributed Step 1: FFTs on the columns (all elements local) Step 2: FFTs on the rows (all elements local) Step 3: FFTs in the Z-dimension (requires communication) Bell, Nishtala, Bonachea, Yelick

41. Performing the 3D FFT (part 3) Can perform Steps 1 and 2 since all the data is available without communication Perform a Global Transpose of the cube –Allows step 3 to continue Transpose Bell, Nishtala, Bonachea, Yelick

42. Communication Strategies for 3D FFT chunk = all rows with same destination pencil = 1 row Three approaches: –Chunk: Wait for 2 nd dim FFTs to finish Minimize # messages –Slab: Wait for chunk of rows destined for 1 proc to finish Overlap with computation –Pencil: Send each row as it completes Maximize overlap and Match natural layout slab = all rows in a single plane with same destination Bell, Nishtala, Bonachea, Yelick

43. NAS FT Variants Performance Summary Slab is always best for MPI; small message cost too high Pencil is always best for UPC; more overlap.5 Tflops Myrinet Infiniband Elan3 Elan3 Elan4 Elan4 #procs 64 256 256 512 256 512 MFlops per Thread Chunk (NAS FT with FFTW) Best MPI (always slabs) Best UPC (always pencils)

44. Bisection Bandwidth Limits Full bisection bandwidth is (too) expensive During an all-to-all communication phase –Effective (per-thread) bandwidth is fractional share –Significantly lower than link bandwidth –Use smaller messages mixed with computation to avoid swamping the network Bell, Nishtala, Bonachea, Yelick

45. Compiler Optimizations Naïve scheme (blocking call for each load/store) not good enough PRE on shared expressions –Reduce the amount of unnecessary communication –Apply also to UPC shared pointer arithmetic Split-phase communication –Hide communication latency through overlapping Message coalescing –Reduce number of messages to save startup overhead and achieve better bandwidth Chen, Iancu, Yelick

46. Benchmarks Gups –Random access (read/modify/write) to distributed array Mcop –Parallel dynamic programming algorithm Sobel –Image filter Psearch –Dynamic load balancing/work stealing Barnes Hut –Shared memory style code from SPLASH2 NAS FT/IS –Bulk communication

47. Performance Improvements % improvement over unoptimized Chen, Iancu, Yelick

48. Data Driven Execution

49. Data-Driven Execution Many algorithms require synchronization with remote processor –Mechanisms: (BUPC extensions) Signaling store: Raise a semaphore upon transfer Remote enqueue: Put a task in a remote queue Remote execution: Floating functions (X10 activities) Many algorithms have irregular data dependencies (LU) –Mechanisms (BUPC extensions) Cooperative multithreading

50. Matrix Factorization Blocks 2D block-cyclic distributed Panel factorizations involve communication for pivoting Completed part of U Completed part of L A(i,j) A(i,k) A(j,i) A(j,k) Trailing matrix to be updated Panel being factored Husbands,Yelick

51. Three Strategies for LU Factorization Husbands,Yelick Organize in bulk-synchronous phases (ScaLAPACK) Factor a block column, then perform updates Relatively easy to understand/debug, but extra synchronization Overlapping phases (HPL): Work associated with on block column factorization can be overlapped Parameter to determine how many (need temp space accordingly) Event-driven multithreaded (UPC Linpack): Each thread runs an event handler loop Tasks: factorization (w/ pivoting), update trailing, update upper Tasks my suspend (voluntarily) to wait for data, synchronization, etc. Data moved with remote gets (synchronization built-in) Must “gang” together for factorizations Scheduling priorities are key to performance and deadlock avoidance

52. UPC-HP Linpack Performance Comparable to HPL (numbers from HPCC database) Faster than ScaLAPACK due to less synchronization Large scaling of UPC code on Itanium/Quadrics (Thunder) 2.2 TFlops on 512p and 4.4 TFlops on 1024p Husbands, Yelick

53. Performance Tuning Iancu, Strohmaier

54. Efficient Use of One-Sided Implementations: need to be efficient and have scalable performance Application level use of NB benefits from new design techniques: finer grained decompositions and overlap Overlap exercises the system in “un-expected” ways Prototyping of implementations for large scale systems is a hard problem: non-linear behavior of networks, communication scheduling is NP-hard Need methodology for fast prototyping: –understand interaction network/CPU at large scale

55. Performance Tuning Performance is determined by overhead, latency and bandwidth, computational characteristics and communication topology It’s all relative: Performance characteristics are determined by system load Basic principles –Minimize communication overhead –Avoid congestion: control injection rate (end-point) avoid hotspots (end-point, network routes) Have to use models. What kind of answers can a model answer?

56. Example: Vector-Add shared double *rdata; double *ldata, *buf; upc_memget(buf, rdata, N); for(i=0; i<N; i++) ldata[i]+= buf[i]; for(i=0; i<N/B; i++) h[i]=upc_memget_nb(buf+i*B,rdata+i*B,B); for(i=0; i<N/B; i++) { sync(h[i]); for(j=0;j<B; j++) ldata[i*B+j]+=buf[i*B+j]; } GET_nb(B 0 ) : GET_nb(B b ) GET_nb(B b+1 ) : GET_nb(B 2b ) sync(B 0 ) compute(B 0 ) : sync(B b ) compute(B b ) GET_nb(B 2b+1 ) : GET_nb(B 3b ) : sync(B N ) compute(B N ) Which implementation is faster? What is B,b? b b b

57. Prototyping Usual approach: use time accurate performance model (applications, automatically tuned collectives) –Models (LogP..) don’t capture important behavior (parallelism, congestion, resource constraints, non-linear behavior) –Exhaustive search of the optimization space –Validated only at low concurrency (~tens of procs), might break at high concurrency, might break for torus networks Our approach: –Use performance model for ideal implementation –Understand hardware resource constraints and the variation of performance parameters (understand trends not absolute values) –Derive implementation constraints to satisfy both optimal implementation and hardware constraints –Force implementation parameters to converge towards optimal

58. Performance Network: bandwidth and overhead Application: communication pattern and schedule (congestion), computation Overhead is determined by message size, communication schedule, hardware flow of control Bandwidth is determined by message size, communication schedule, fairness of allocation Iancu, Strohmaier

59. Understand network behavior in the presence of non-blocking communication (Infiniband, Elan) Develop performance model for scenarios widely encountered in applications ( p2p, scatter, gather, all-to-all ) and a variety of aggressive optimization techniques ( strip mining, pipelining, communication schedule skewing ) Use both micro-benchmarks and application kernels to validate approach Validation Iancu, Strohmaier

60. Findings Can choose optimal values for implementation parameters Time accurate model for an implementation: hard to develop, inaccurate at high concurrency Methodology does not require exhaustive search of the optimization space ( only p2p and qualitative behavior of gather ) In practice one can produce templatized implementations for an algorithm and use our approach to determine optimal values: code generation (UPC), automatic tuning of collective operations, application development Need to further understand the mathematical and statistical properties

61. End

62. Avoid synchronization: Data-driven Execution Many algorithms require synchronization with remote processor What is the right mechanism in a PGAS model for doing this? Is it still one-sided? Part 3: Event-Driven Execution Models

63. Mechanisms for Event-Driven Execution “Put” operation does a send side notification –Needed for memory consistency model ordering Need to signal remote side on completion Strategies: –Have remote side do a “get” (works in some algorithms) –Put + strict flag write: do a put, wait for completion, then do another (strict) put –Pipelined put + put-flag: works only on ordered networks –Signaling put: add new “store” operation that embeds signal (2 nd remote address) into single message

64. Mechanisms for Event-Driven Execution Preliminary results on Opteron + Infiniband