Feb. 19, 2008 Multicore Processor Technology and Managing Contention for Shared Resource Cong Zhao Yixing Li

Moore’s Law Transistors per integrated circuit would double every 2 years. Power requirements in relation to transistor size would double in 1-2 years. Density at minimum cost per transistor, and so on…… Integrated circuits would double in performance every 18 months(By David House) For Single-Core, the only way to improve the performance is increasing the clock frequency. 2

Why Multi-Core? There are many reasons : Difficult to make single-core clock frequencies even higher Deeply pipelined circuits: – heat problems – speed of light problems – difficult design and verification – large design teams necessary Many new applications are multithreaded 3

Why Multi-Core? 4 But the leading reason is: To continue the raw performance growth that customers have come to expect from Moore’s law scaling without being overwhelmed by the growth in power consumption. As single core designs were pushed to ever higher clock speeds, the power required grew at a faster rate than the frequency, and lead to designs that were complex, power hungry, and unmanageable !!!

Why Multi-Core? 5

What is Multi-Core? A single computing component with two or more independent actual central processing units(called "cores") Attributes:  Application Class  Power/Performance  Processing Elements  Memory System  Etc. 6

Application Class  There are two broad classes of processing into which an application can fall: data processing dominated and control dominated.  Data Processing Dominated  A sequence of operations on a stream of data with little or no data reuse.  Image processing, audio processing, and wireless baseband processing.  Control Dominated  Often need to keep track of large amounts of state and often have a high amount of data reuse.  file compression/decompression, network processing. 7

Power/Performance  In the past decade, power has joined performance as a first class design. Many applications and devices have strict performance and power requirements.  Mobile phone, Laptop, Server  “cloud” computing (warehousescale computers) 8 These cloud computing centers are now consuming more energy than heavy manufacturing in the United States.

Processing Elements Architecture and Microarchitecture Architecture:  Instruction set architecture (ISA), defines the hardware softwareinterface.  Reduced instruction set computer(RISC)  Complex instruction set computer (CISC). Microarchitecture:  The microarchitecture is the implementation of the ISA.  In-order processing element.  Out-of-order processing element.  SIMD, VLIW 9

Memory System  In uniprocessor designs, the memory system was a rather simple component, consisting of a few levels of cache to feed the single processor with data and instructions.  In Multi-core design:  Consistency model.  Cache configuration.  Cache coherence support.  Intrachip interconnect.  All of these determine how cores communicate impacting programmability, parallel application performance, and the number of cores that the system can adequately support. 10

Memory System  Consistency Model  Strong Consistency and Weak Consistency 11 0

Memory System  Cache Configuration  Caches give processing elements a fast, high bandwidth local memory to work with.  Caches can be tagged and managed automatically by hardware or be explicitly managed local store memory.  The amount of cache re quired is very application dependent.  The first level of cache (L1)is usually rather small, fast, and private to each processing element. Subsequent levels (L2)can be larger, slower, and shared among processing elements. 12

Memory System  Intrachip interconnect.  Bus, Crossbar, Ring, and Network-on-chip (NoC)  Cache coherence  Broadcast based and Directory based. 13

14 Architecture of A Multicore System Fig2-1. Schematic of a Multicore System with Two Memory Domains Compete for the shared resources! Problems?

15 Cache Contention  Thread A request a line not in the cache (a cache miss) and the cache is full  Some data must be evicted to free up a line  The evicted line might belong to B or A itself Hurt the performance Thread A Thread B

16 Cache Miss Frequency Temporal locality Reuse frequency Miss frequency A.McfRather poorLowHigh B. PovrayExcellentHigh →0→0 C. MilcPoorVery lowVery high Fig2-2. Example Memory-Reuse Profiles from SPEC CPU2006 Suite

17 Pain Metrix  Pain(A|B) = SA * ZB  Pain(B|A) = SB * ZA  Pain(A,B) = Pain(A|B) + Pain(B|A) Performance degradation of A when A runs with B relative to running solo S Sensitivity, measures how much a thread suffers whenever it shares the cache with other threads Z Intensity, measures how much a thread hurts other threads

18 Evaluation of Pain Model

19 Evaluation of Pain Model Fig2-3. Performance of Estimated Best Schedules Compared with Actual Best Schedules

20 Evaluation of Pain Model Fig2-4. Worst-Case Performance under DIO Relative to the Default Linux Scheduler ·Average performance improvement: 11% ·High-miss-rate applications must be kept apart

21 Problem of Pain Model  2 memory domain  2 cores per domain  8 memory domain  2 cores per domain What about? Example shows..

22 Future Work of Pain Model Combination of schedules 2 memory domain 2 cores per domain 3 4 memory domain 2 cores per domain 105  High-miss-rate applications must be kept apart  We can combine 1 of 4 high-miss-rate cores with 1 of 4 low- miss-rate cores within one domain  Combination of schedules can be reduced to 24

23 Conclusions  The main advantage to multicore systems is that raw performance increase can come from increasing the number of cores rather than frequency, which translates into a slower growth in power consumption. However, this approach represents a significant gamble because parallel programming science has not advanced nearly as fast as our ability to build parallel hardware.

24 Thank you