Multicore/Manycore Processors - St. Francis Xavier University

Report
Joram Benham
April 2, 2012



Introduction
Motivation
Multicore Processors
 Overview, CELL

Advantages of CMPs
 Throughput, Latency


Challenges
Future of Multicore

Multicore processors
 Several/many cores on the same chip
 Dual/quad core – two/four cores

AKA Chip-multiprocessors (CMPs)

Instruction-Level Parallelism
 Pipelining – split execution into stages
 Superscalar – issue multiple instruction each cycle
 Out-of-order execution
 Branch prediction

Take advantage of implicit program
parallelism – instruction independence
1.
2.
3.
Limited amount of implicit parallelism in
sequentially designed/coded programs
Circuitry for pipelining becomes complex
after 10-20 stages
Power – circuitry for ILP exploitation results
in exponentially more power being used
Intel processor power over time. Power in Watts on y-axis, years on x-axis.
AKA Multicore/Manycore Processor


Getting harder to build better uniprocessors
CMPs are less difficult
 Can reuse/modify old designs
 Add modified copies to same chip

Requires a paradigm shift
 From Von Neumann model to parallel
programming model
 Thread-level parallelism + instruction-level
parallelism




CELL CMP – heterogeneous
Developed by Sony, Toshiba, IBM
Built for Sony’s PlayStation 3
Contains 9 cores
 1 Power Processing Element (PPE)
 8 Synergistic Processing Elements (SPEs)
Throughput, Latency

Web-server throughput
 Handle many independent service requests



Collections of uniprocessor servers used
Then, multiprocessor systems
CMP approach
 Use less power for communication
 Reducing clock-speeds

General rule:
 “The simpler the pipeline, the lower the power.”
 Simple cores – less power used
 Less speed, but more cores available to handle
requests
Comparison of power usage by equivalent narrow issue/in-order processors, and
wide-issue/out-of-order processors on throughput-oriented software.

Server applications:
 High thread-level parallelism
 Lower instruction-level parallelism, high cache
miss rates
 Results in idle processor time on uniprocessors

Hardware multithreading
 Coarse-grained: stalls trigger switches
 Fine-grained: switch threads continuously
 Simultaneous: Run multiple threads using
superscalar issuing


More cores = higher total hardware thread
count
What kind of cores should be added?
 Fewer larger, more complex cores
▪ Individual threads complete faster
 Many smaller, simpler cores
▪ Slightly slower – but more cores means more threads,
and higher throughput

Latency is more important in some programs
 E.g. Desktop applications, compilation


CMPs are closer together on chip – less
communication time
Two ways CMPs help with latency
 Parallelize the code for responsive applications
 Run sequential applications on their own
hardware threads – no competition between
threads
Power and Temperature, Cache Coherence, Memory Access, Paradigm
Shift, Starvation


In theory: two cores on the same chip = twice
as much power + lots of heat
Solutions:
 Reduce core clock speeds
 Implement a power control unit
CELL chip-multiprocessor thermal diagram.

Multiple cores, independent local caches
 Load same block of main memory into cache –
may result in data inconsistency

Cache coherence schemes
 Snooping: Watch the communication bus
 Directory-based: Keep track of which memory
locations are being shared in multiple caches

We need more memory to share among
multicore processors
 64-bit processors – helps address the issue: more
addressable memory
 Useless if we cannot access it quickly
 Disk speed slows everyone down
“To use multicore, you really have to use multiple
threads. If you know how to do it, it's not bad. But the
first time you do it there are lots of ways to shoot
yourself in the foot. The bugs you introduce with
multithreading are so much harder to find.”
 Have to educate programmers
 Convince them to make their programs concurrent


Sequential programs will not use all cores
 Some cores “starve”

Shared cache usage
 One core evicts another core’s data
 Other core has to keep accessing main memory
Multicore, Manycore, Hybrids



Instruction-level parallelism reaching its
limits
CMPs help with throughput and latency
Two types of CMP will emerge
 “Manycore”: large number of small, simple cores,
targets at servers/throughput
 “Multicore”: fewer, faster superscalar cores for
very latency sensitive programs
 “Hybrids”: heterogeneous combinations
Hammond, L., Laudon, J., Olukotun, K. Chip Multiprocessor Architecture: Techniques
to Improve Throughput and Latency. Morgan and Claypool, 2007.
Hennessy, J. L., Patterson, D. A. Computer Architecture: A Quantitative Approach.
San Francisco: Morgan Kaufmann Publishers, 2007.
Mashiyat, A. S. “Multi/Many Core Systems.” St. Francis Xavier University course
presentation, 2011.
Schauer, Bryan. “Multicore Processors – A Necessity.” Proquest Discovery Guides.
September 2008. Web. Accessed April 2 2012.
<http://www.csa.com/discoveryguides/multicore/review.pdf>

similar documents