1. High Performance Computing1 High Performance Computing (CS 680) Lecture 2a: Overview of High Performance Processors * Jeremy R. Johnson *This lecture was derived from material in the text (HPC Chap. 1-2).

2. High Performance Computing2 Introduction Objective: To review recent developments in the design of high performance microprocessors. To indicate how these features effect program performance. An example program will be used to illustrate benchmarking techniques and the effect of compiler optimizations and code organization on performance. We will indicate how changes in software can improve performance by better utilizing the underlying hardware. Our goal for the course is to understand this behavior. Topics –pipelining –instruction level parallelism, superscalar and out of order execution –Memory Hierarchy: cache, virtual memory

3. High Performance Computing3 RISC vs. CISC CISC: instruction set made up of powerful instructions close to primitives in a high-level language such as C or FORTRAN RISC: low level instructions are emphasized. RISC is a label most commonly used for a set of instruction set architecture characteristics chosen to ease the use of aggressive implementation techniques found in high-performance processors (John Mashey) Prevalence began in mid-1980s (earlier example CDC 6600) when more transistors and better compilers became available. Trade complex instructions for faster clock rate and more room for extra registers, cache and advanced performance techniques.

4. High Performance Computing4 Characterizing RISC Instruction pipelining Pipelining floating point execution Uniform instruction length Delayed branching Load/Store architecture Simple addressing modes

5. High Performance Computing5 Pipelining Instruction pipelining –Instruction Fetch –Instruction Decode –Operand Fetch –Execute –Writeback IFIDFEW IFIDFEW IFIDFEW

6. High Performance Computing6 Branches and Hazards If a branch is executed the pipeline may need to be flushed since the wrong instructions may have been started. IFIDFEW IFIDFEW IFIDFEW IFIDFEW IFIDFE guess sure

7. High Performance Computing7 Advanced Techniques Superscalar Processors –issue more than one instruction per cycle –can’t have dependencies or hardware conflict –for example can execute an add simultaneously with a mult Superpipeling –more stages in the pipeline Out of order and speculative execution –maintain semantics but allow instructions to be computed in different order –may need to guess which instruction to execute –depends on difference between computation and execution

8. High Performance Computing8 Post-RISC Pipeline IFID IRB E RR R Instruction Reorder Buffer Rename Registers Branch Prediction

9. High Performance Computing9 Memory Hierarchy SRAM vs. DRAM –small fast memory vs. large slow memory –principle of locality Registers Cache (level 1) Cache (level 2) Main memory Disk

10. High Performance Computing10 Memory Access Speed on DEC 21164 Alpha Clock Speed 500 MHz (= 2 ns clock rate) Registers (2 ns) L1 On-Chip (4 ns) L2 On-Chip (5 ns) L3 Off-Chip (30 ns) Memory (220 ns)

11. High Performance Computing11 Common Framework for Memory Hierarchies Question 1: Where can a block be placed? –One place (direct mapped), a few places (set associative), or any place (fully associative) Question 2: How is a block found? –There are four methods: indexing, limited search, full search, or table lookup Question 3: Which block should be replaced on a cache miss? –Typically least recently used or random block Question 4: What happens on writes? –Write-through or write-back

12. High Performance Computing12 Mapping to Cache Cache - the level of the memory hierarchy between the CPU and main memory. Direct-Mapped Cache - memory mapped to one location in cache (Block address) mod (Number of block in cache) Number of blocks is typically a power of two  cache location obtained from low-order bits of address.... 0 1 2 3 28 29 30 31 0 1 2 3

13. High Performance Computing13 Locating an Data in the Cache Compare cache index (mapping) to Tag (high- order bits) to see if element is currently in cache Valid bit used to indicate whether data in cache is valid A hit occurs when the data is in cache, otherwise it is a miss The extra time required when a cache miss occurs is called the miss penalty

14. High Performance Computing14 Example 32-word memory 8-word cache

15. High Performance Computing15 Cache Organization Since cache is smaller than memory more than one address must map to same line in cache Direct-Mapped Cache –address mod cache size (only one location when memory address gets mapped to) Fully Associative Cache –address can be mapped anywhere in cache –need tag and associative search to find if element in cache Set-Associative Cache –compromise between two extremes –element can map to several locations

16. High Performance Computing16 Model for Cache Misses Compulsory misses –These are cache misses caused by the first access to a block that has never been in the cache Capacity misses (cold-start) –These are cache misses caused when the cache cannot contain all the blocks needed during execution of a program. Conflict misses (collision) –These are cache misses that occur in a set associative or direct mapped cache when multiple blocks compete for the same set. These misses are eliminated with a fully associative cache.

17. High Performance Computing17 Measuring Cache Performance CPU time = (CPU execution clock cycles + Memory stall clock cycles)  Clock-cycle time Memory stall clock cycles = Read-stall cycles + Write-stall cycles Read-stall cycles = Reads/program  Read miss rate  Read miss penalty Write-stall cycles = (Writes/program  Write miss rate  Write miss penalty) + Write buffer stalls (assumes write-through cache) Write buffer stalls should be negligible and write and read miss penalties equal (cost to fetch block from memory) Memory stall clock cycles = Mem access/program  miss rate  miss penalty

18. High Performance Computing18 Virtual Memory Decouple physical addresses (memory locations) from addresses used by a program. Programmer sees a large memory with the same virtual addresses independent of where the program is actually placed in memory. –Virtual to physical mapping performed via a page table –Since page tables can be in virtual memory, there could be several table lookups for a single memory reference. –TLB (translation lookaside buffer) is a cache to store commonly used virtual to physical maps. Page Fault –when page is not in memory it must be brought in (from disk) –very slow (usually occurs with OS intervention)

19. High Performance Computing19 Improving Memory Performance Larger and wider caches Cache bypass Interleaved and pipelined memory systems Prefetching Post-RISC effects on memory New memory trends