Lecture 29 Computer Arch. From: Supercomputing in Plain English Part VII: Multicore Madness Henry Neeman, Director OU Supercomputing Center for Education & Research University of Oklahoma Wednesday October 17 2007 Outline The March of Progress Multicore/Many-core Basics Software Strategies for Multicore/Many-core A Concrete Example: Weather Forecasting Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 2 The March of Progress OU’s TeraFLOP Cluster, 2002 10 racks @ 1000 lbs per rack 270 Pentium4 Xeon CPUs, 2.0 GHz, 512 KB L2 cache 270 GB RAM, 400 MHz FSB 8 TB disk Myrinet2000 Interconnect 100 Mbps Ethernet Interconnect OS: Red Hat Linux Peak speed: 1.08 TFLOP/s (1.08 trillion calculations per second) One of the first Pentium4 clusters! boomer.oscer.ou.edu Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 4 TeraFLOP, Prototype 2006, Sale 2011 9 years from room to chip! http://news.com.com/2300-1006_3-6119652.html Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 5 Moore’s Law In 1965, Gordon Moore was an engineer at Fairchild Semiconductor. He noticed that the number of transistors that could be squeezed onto a chip was doubling about every 18 months. It turns out that computer speed is roughly proportional to the number of transistors per unit area. Moore wrote a paper about this concept, which became known as “Moore’s Law.” Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 6 log(Speed) Moore’s Law in Practice Year Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 7 log(Speed) Moore’s Law in Practice Year Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 8 log(Speed) Moore’s Law in Practice Year Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 9 log(Speed) Moore’s Law in Practice Year Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 10 log(Speed) Moore’s Law in Practice Year Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 11 Fastest Supercomputer vs. Moore Fastest Supercomputer in the World 1000000 www.top500.org Speed in GFLOP/s 100000 10000 Fastest Moore 1000 100 10 1 1992 1994 1996 1998 2000 2002 2004 2006 2008 Year Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 12 The Tyranny of the Storage Hierarchy The Storage Hierarchy [5] Fast, expensive, few Registers Cache memory Main memory (RAM) Hard disk Removable media (e.g., DVD) Internet Slow, cheap, a lot [6] Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 14 RAM is Slow The speed of data transfer between Main Memory and the CPU is much slower than the speed of calculating, so the CPU spends most of its time waiting for data to come in or go out. CPU 351 GB/sec[7] Bottleneck 10.66 GB/sec[9] (3%) Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 15 Why Have Cache? Cache is nearly the same speed as the CPU, so the CPU doesn’t have to wait nearly as long for stuff that’s already in cache: it can do more operations per second! CPU 351 GB/sec[7] 253 GB/sec[8] (72%) 10.66 GB/sec[9] (3%) Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 16 Storage Use Strategies Register reuse: Do a lot of work on the same data before working on new data. Cache reuse: The program is much more efficient if all of the data and instructions fit in cache; if not, try to use what’s in cache a lot before using anything that isn’t in cache. Data locality: Try to access data that are near each other in memory before data that are far. I/O efficiency: Do a bunch of I/O all at once rather than a little bit at a time; don’t mix calculations and I/O. Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 17 A Concrete Example OSCER’s big cluster, topdawg, has Irwindale CPUs: single core, 3.2 GHz, 800 MHz Front Side Bus. The theoretical peak CPU speed is 6.4 GFLOPs (double precision) per CPU, and in practice we’ve gotten as high as 94% of that. So, in theory each CPU could consume 143 GB/sec. The theoretical peak RAM bandwidth is 6.4 GB/sec, but in practice we get about half that. So, any code that does less than 45 calculations per byte transferred between RAM and cache has speed limited by RAM bandwidth. Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 18 Good Cache Reuse Example A Sample Application Matrix-Matrix Multiply Let A, B and C be matrices of sizes nr nc, nr nk and nk nc, respectively: a1,1 a 2,1 A a3,1 anr ,1 a1, 2 a2 , 2 a3, 2 anr , 2 a1,3 a2 , 3 a3,3 anr ,3 a1,nc a2,nc a3,nc anr ,nc b1,1 b1, 2 b 2,1 b2, 2 B b3,1 b3, 2 bnr ,1 bnr , 2 b1,3 b2,3 b3,3 bnr ,3 b1,nk b2,nk b3,nk bnr ,nk c1,1 c1, 2 c 2,1 c2, 2 C c3,1 c3, 2 cnk ,1 cnk , 2 c1,3 c2 , 3 c3,3 cnk ,3 c1,nc c2,nc c3,nc cnk ,nc The definition of A = B • C is nk ar ,c br ,k ck ,c br ,1 c1,c br , 2 c2,c br ,3 c3,c br ,nk cnk ,c k 1 for r {1, nr}, c {1, nc}. Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 20 Matrix Multiply: Naïve Version SUBROUTINE matrix_matrix_mult_naive & IMPLICIT NONE INTEGER,INTENT(IN) :: nr, nc, nq REAL,DIMENSION(nr,nc),INTENT(OUT) REAL,DIMENSION(nr,nq),INTENT(IN) REAL,DIMENSION(nq,nc),INTENT(IN) (dst, src1, src2, & nr, nc, nq) :: dst :: src1 :: src2 INTEGER :: r, c, q DO c = 1, nc DO r = 1, nr dst(r,c) = 0.0 DO q = 1, nq dst(r,c) = dst(r,c) + src1(r,q) * src2(q,c) END DO END DO END DO END SUBROUTINE matrix_matrix_mult_naive Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 21 Performance of Matrix Multiply 800 Matrix-Matrix Multiply 700 CPU sec 600 500 400 300 Better Init 200 100 0 0 10000000 20000000 30000000 40000000 50000000 60000000 Total Problem Size in bytes (nr*nc+nr*nq+nq*nc) Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 22 Tiling Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 23 Tiling Tile: A small rectangular subdomain of a problem domain. Sometimes called a block or a chunk. Tiling: Breaking the domain into tiles. Tiling strategy: Operate on each tile to completion, then move to the next tile. Tile size can be set at runtime, according to what’s best for the machine that you’re running on. Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 24 Tiling Code SUBROUTINE matrix_matrix_mult_by_tiling (dst, src1, src2, nr, nc, nq, & & rtilesize, ctilesize, qtilesize) IMPLICIT NONE INTEGER,INTENT(IN) :: nr, nc, nq REAL,DIMENSION(nr,nc),INTENT(OUT) :: dst REAL,DIMENSION(nr,nq),INTENT(IN) :: src1 REAL,DIMENSION(nq,nc),INTENT(IN) :: src2 INTEGER,INTENT(IN) :: rtilesize, ctilesize, qtilesize INTEGER :: rstart, rend, cstart, cend, qstart, qend DO cstart = 1, nc, ctilesize cend = cstart + ctilesize - 1 IF (cend > nc) cend = nc DO rstart = 1, nr, rtilesize rend = rstart + rtilesize - 1 IF (rend > nr) rend = nr DO qstart = 1, nq, qtilesize qend = qstart + qtilesize - 1 IF (qend > nq) qend = nq CALL matrix_matrix_mult_tile(dst, src1, src2, nr, nc, nq, & & rstart, rend, cstart, cend, qstart, qend) END DO END DO END DO END SUBROUTINE matrix_matrix_mult_by_tiling Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 25 Multiplying Within a Tile SUBROUTINE matrix_matrix_mult_tile (dst, src1, src2, nr, nc, nq, & & rstart, rend, cstart, cend, qstart, qend) IMPLICIT NONE INTEGER,INTENT(IN) :: nr, nc, nq REAL,DIMENSION(nr,nc),INTENT(OUT) :: dst REAL,DIMENSION(nr,nq),INTENT(IN) :: src1 REAL,DIMENSION(nq,nc),INTENT(IN) :: src2 INTEGER,INTENT(IN) :: rstart, rend, cstart, cend, qstart, qend INTEGER :: r, c, q DO c = cstart, cend DO r = rstart, rend IF (qstart == 1) dst(r,c) = 0.0 DO q = qstart, qend dst(r,c) = dst(r,c) + src1(r,q) * src2(q,c) END DO END DO END DO END SUBROUTINE matrix_matrix_mult_tile Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 26 Reminder: Naïve Version, Again SUBROUTINE matrix_matrix_mult_naive & IMPLICIT NONE INTEGER,INTENT(IN) :: nr, nc, nq REAL,DIMENSION(nr,nc),INTENT(OUT) REAL,DIMENSION(nr,nq),INTENT(IN) REAL,DIMENSION(nq,nc),INTENT(IN) (dst, src1, src2, & nr, nc, nq) :: dst :: src1 :: src2 INTEGER :: r, c, q DO c = 1, nc DO r = 1, nr dst(r,c) = 0.0 DO q = 1, nq dst(r,c) = dst(r,c) + src1(r,q) * src2(q,c) END DO END DO END DO END SUBROUTINE matrix_matrix_mult_naive Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 27 Performance with Tiling Matrix-Matrix Mutiply Via Tiling Matrix-Matrix Mutiply Via Tiling (log-log) 250 1000 200 100 512x256 10 CPU sec 150 512x512 1024x512 100 1024x1024 1 50 1E+08 10000000 1000000 100000 10000 1000 100 10 2048x1024 Better 0.1 0 100000000 10000000 1000000 100000 10000 1000 100 10 Tile Size (bytes) Tile Size (bytes) Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 28 The Advantages of Tiling It allows your code to exploit data locality better, to get much more cache reuse: your code runs faster! It’s a relatively modest amount of extra coding (typically a few wrapper functions and some changes to loop bounds). If you don’t need tiling – because of the hardware, the compiler or the problem size – then you can turn it off by simply setting the tile size equal to the problem size. Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 29 Why Does Tiling Work Here? Cache optimization works best when the number of calculations per byte is large. For example, with matrix-matrix multiply on an n × n matrix, there are O(n3) calculations (on the order of n3), but only O(n2) bytes of data. So, for large n, there are a huge number of calculations per byte transferred between RAM and cache. Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 30 Multicore/Many-core Basics What is Multicore? In the olden days (i.e., the first half of 2005), each CPU chip had one “brain” in it. More recently, each CPU chip has 2 cores (brains), and, starting in late 2006, 4 cores. Jargon: Each CPU chip plugs into a socket, so these days, to avoid confusion, people refer to sockets and cores, rather than CPUs or processors. Each core is just like a full blown CPU, except that it shares its socket with one or more other cores – and therefore shares its bandwidth to RAM. Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 32 Dual Core Core Core Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 33 Quad Core Core Core Core Core Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 34 Oct Core Core Core Core Core Core Core Core Core Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 35 The Challenge of Multicore: RAM Each socket has access to a certain amount of RAM, at a fixed RAM bandwidth per SOCKET. As the number of cores per socket increases, the contention for RAM bandwidth increases too. At 2 cores in a socket, this problem isn’t too bad. But at 16 or 32 or 80 cores, it’s a huge problem. So, applications that are cache optimized will get big speedups. But, applications whose performance is limited by RAM bandwidth are going to speed up only as fast as RAM bandwidth speeds up. RAM bandwidth speeds up much slower than CPU speeds up. Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 36 The Challenge of Multicore: Network Each node has access to a certain number of network ports, at a fixed number of network ports per NODE. As the number of cores per node increases, the contention for network ports increases too. At 2 cores in a socket, this problem isn’t too bad. But at 16 or 32 or 80 cores, it’s a huge problem. So, applications that do minimal communication will get big speedups. But, applications whose performance is limited by the number of MPI messages are going to speed up very very little – and may even crash the node. Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 37 Multicore/Many-core Problem Most multicore chip families have relatively small cache per core (e.g., 2 MB) – and this problem seems likely to remain. Small TLBs make the problem worse: 512 KB per core rather than 2 MB. So, to get good cache reuse, you need to partition algorithm so subproblem needs no more than 512 KB. Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 38 The T.L.B. on a Current Chip On Intel Core Duo (“Yonah”): Cache size is 2 MB per core. Page size is 4 KB. A core’s data TLB size is 128 page table entries. Therefore, D-TLB only covers 512 KB of cache. Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 39 The T.L.B. on a Current Chip On Intel Core Duo (“Yonah”): Cache size is 2 MB per core. Page size is 4 KB. A core’s data TLB size is 128 page table entries. Therefore, D-TLB only covers 512 KB of cache. The cost of a TLB miss is 49 cycles, equivalent to as many as 196 calculations! (4 FLOPs per cycle) http://www.digit-life.com/articles2/cpu/rmma-via-c7.html Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 40 What Do We Need? We need much bigger caches! TLB must be big enough to cover the entire cache. It’d be nice to have RAM speed increase as fast as core counts increase, but let’s not kid ourselves. Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 41 To Learn More Supercomputing http://www.oscer.ou.edu/education.php Supercomputing in Plain English: Multicore Madness Wednesday October 17 2007 42 Computer Architecture Fall 2007 Lecture 29. CMPs & SMTs Adapted from Mary Jane Irwin ( www.cse.psu.edu/~mji ) [Adapted from Computer Organization and Design, Patterson & Hennessy, © 2005] Lecture 29 Fall Multithreading on A Chip Find a way to “hide” true data dependency stalls, cache miss stalls, and branch stalls by finding instructions (from other process threads) that are independent of those stalling instructions Multithreading – increase the utilization of resources on a chip by allowing multiple processes (threads) to share the functional units of a single processor Lecture 29 Processor must duplicate the state hardware for each thread – a separate register file, PC, instruction buffer, and store buffer for each thread The caches, TLBs, BHT, BTB can be shared (although the miss rates may increase if they are not sized accordingly) The memory can be shared through virtual memory mechanisms Hardware must support efficient thread context switching Fall Types of Multithreading on a Chip Fine-grain – switch threads on every instruction issue Round-robin thread interleaving (skipping stalled threads) Processor must be able to switch threads on every clock cycle Advantage – can hide throughput losses that come from both short and long stalls Disadvantage – slows down the execution of an individual thread since a thread that is ready to execute without stalls is delayed by instructions from other threads Coarse-grain – switches threads only on costly stalls (e.g., L2 cache misses) Advantages – thread switching doesn’t have to be essentially free and much less likely to slow down the execution of an individual thread Disadvantage – limited, due to pipeline start-up costs, in its ability to overcome throughput loss - Pipeline must be flushed and refilled on thread switches Lecture 29 Fall Multithreaded Example: Sun’s Niagara (UltraSparc T1) 1.2 GHz 1.0 GHz Cache (I/D/L2) 32K/64K/ (8M external) 16K/8K/3M Issue rate 4 issue 1 issue Pipe stages 14 stages 6 stages BHT entries 16K x 2-b None TLB entries 128I/512D 64I/64D Memory BW 2.4 GB/s ~20GB/s Transistors 29 million 200 million Power (max) 53 W Lecture 29 <60 W 4-way MT SPARC pipe Clock rate 4-way MT SPARC pipe 64-b 4-way MT SPARC pipe 64-b 4-way MT SPARC pipe Data width 4-way MT SPARC pipe Niagara 4-way MT SPARC pipe Ultra III 4-way MT SPARC pipe Eight fine grain multithreaded single-issue, in-order cores (no speculation, no dynamic branch prediction) 4-way MT SPARC pipe Crossbar I/O shared funct’s 4-way banked L2$ Memory controllers Fall Niagara Integer Pipeline Cores are simple (single-issue, 6 stage, no branch prediction), small, and power-efficient Fetch Thrd Sel Decode RegFile x4 I$ Inst bufx4 ITLB Thrd Sel Mux Thrd Sel Mux Decode Thread Select Logic Execute ALU Mul Shft Div Memory D$ DTLB Stbufx4 WB Crossbar Interface Instr type Cache misses Traps & interrupts Resource conflicts PC logicx4 From MPR, Vol. 18, #9, Sept. 2004 Lecture 29 Fall Simultaneous Multithreading (SMT) A variation on multithreading that uses the resources of a multiple-issue, dynamically scheduled processor (superscalar) to exploit both program ILP and threadlevel parallelism (TLP) Most SS processors have more machine level parallelism than most programs can effectively use (i.e., than have ILP) With register renaming and dynamic scheduling, multiple instructions from independent threads can be issued without regard to dependencies among them - Need separate rename tables (ROBs) for each thread - Need the capability to commit from multiple threads (i.e., from multiple ROBs) in one cycle Intel’s Pentium 4 SMT called hyperthreading Lecture 29 Supports just two threads (doubles the architecture state) Fall Threading on a 4-way SS Processor Example Coarse MT Fine MT SMT Issue slots → Thread A Thread B Time → Thread C Thread D Lecture 29 Fall Multicore Xbox360 – “Xenon” processor Lecture 29 To provide game developers with a balanced and powerful platform Three SMT processors, 32KB L1 D$ & I$, 1MB UL2 cache 165M transistors total 3.2 Ghz Near-POWER ISA 2-issue, 21 stage pipeline, with 128 128-bit registers Weak branch prediction – supported by software hinting In order instructions Narrow cores – 2 INT units, 2 128-bit VMX units, 1 of anything else An ATI-designed 500MZ GPU w/ 512MB of DDR3DRAM 337M transistors, 10MB framebuffer 48 pixel shader cores, each with 4 ALUs Fall Xenon Diagram Core 1 Core 2 L1D L1I L1D L1I L1D L1I XMA Dec Core 0 1MB UL2 MC1 MC0 512MB DRAM BIU/IO Intf Lecture 29 SMC GPU DVD HDD Port Front USBs (2) Wireless MU ports (2 USBs) Rear USB (1) Ethernet IR Audio Out Flash Systems Control 3D Core 10MB EDRAM Video Out Analog Chip Video Out Fall The PS3 “Cell” Processor Architecture Composed of a Non-SMP Architecture 234M transistors @ 4Ghz 1 Power Processing Element, 8 “Synergistic” (SIMD) PE’s 512KB L2 $ - Massively high bandwidth (200GB/s) bus connects it to everything else The PPE is strangely similar to one of the Xenon cores - Almost identical, really. Slight ISA differences, and fine-grained MT instead of real SMT The real differences lie in the SPEs (21M transistors each) - An attempt to ‘fix’ the memory latency problem by giving each processor complete control over it’s own 256KB “scratchpad” – 14M transistors – Direct mapped for low latency - 4 vector units per SPE, 1 of everything else – 7M trans. Lecture 29 Fall The PS3 “Cell” Processor Architecture Lecture 29 Fall How to make use of the SPEs Lecture 29 Fall What about the Software? Lecture 29 Makes use of special IBM “Hypervisor” Like an OS for OS’s Runs both a real time OS (for sound) and non-real time (for things like AI) Software must be specially coded to run well The single PPE will be quickly bogged down Must make use of SPEs wherever possible This isn’t easy, by any standard What about Microsoft? Development suite identifies which 6 threads you’re expected to run Four of them are DirectX based, and handled by the OS Only need to write two threads, functionally http://ps3forums.com/showthread.php?t=22858 Fall Next Lecture and Reminders Reminders Lecture 29 Final is Thursday, December 13 from 10-11:50 AM in ITT 322 Fall