Systems II 15-319, spring 2010 6th Lecture, Jan 28th Lecture MotivationPerformance BottlenecksPerformance BottlenecksPerformance BottlenecksSystem Design Performance ConsiderationsSystem Design Performance Considerations# Blade Centers# Blades/Blade Center# Processors/Blade # Cores/ProcessorWhy a Multi-core Processor?Why a Multi-core Processor?Multicore processorsDesign Concerns: processor frequencyCore Design (1/4)Core Design (2/4)Core Design (3/4)Core Design (4/4)Design ConcernsDesign ConcernsExample: Given This Data…..Example: Given This Data…..Example: Calculate Execution Time for a.exeExample: Calculate Execution Time for a.exeExample: Calculate Execution Time for a.exeExample: Calculate Execution Time for a.exeExample: Calculate Execution Time for a.exeExample: Calculate Execution Time for a.exeDesign ConcernsTowards Faster NetworksHigh BW & Speed NetworksGigabit EthernetMyrinetInfinibandCarnegie MellonSpring 2010 ©15-319 Introduction to Cloud ComputingIntroduction to Cloud ComputingMajd F. SakrSystems II 15‐319, spring 2010 6th Lecture, Jan 28thCarnegie MellonSpring 2010 ©15-319 Introduction to Cloud Computing2Lecture Motivation System Impact on Performance System Considerations Performance EvaluationCarnegie MellonSpring 2010 ©15-319 Introduction to Cloud ComputingPerformance Bottlenecks Consider bandwidth and latency between these layersL3MemoryQuad Core ProcessorDiskcorecorecorecoreL2Quad Core ProcessorcorecorecorecoreL2L1L1Carnegie MellonSpring 2010 ©15-319 Introduction to Cloud ComputingPerformance Bottlenecks Consider bandwidth and latency of all layersHow about pipes of this network??Carnegie MellonSpring 2010 ©15-319 Introduction to Cloud ComputingPerformance Bottlenecks Consider bandwidth and latency of all of these pipesHow about a network between different blade centers??Carnegie MellonSpring 2010 ©15-319 Introduction to Cloud ComputingSystem Design Performance ConsiderationsWhere does your application live? across….Blade CentersBladesProcessorsCores++Latency , Storage, Processors, SizeBandwidthCarnegie MellonSpring 2010 ©15-319 Introduction to Cloud Computing Number of blade Centers Number of blades in each Blade Center Number of processors in each Blade or Motherboard Number of cores in each processor Processor Frequency Per core: Order of superscalars Number of Registers Bandwidth between register  File and the Functional Unit Memory design Bandwidth between all these componentsSystem Design Performance ConsiderationsCarnegie MellonSpring 2010 ©15-319 Introduction to Cloud Computing# Blade Centers A blade center OR group of blade centers could be a “cloud” This # affects the blade centers’ networking issues: More bandwidth & less latency means “better communication” If the blade centers are connected to form a “cloud” for intensive computation: Processes are tightly coupled: There is a lot of communication to achieve cooperative work Maximizing network performance would maximize the cloud efficiency and productivityCarnegie MellonSpring 2010 ©15-319 Introduction to Cloud Computing# Blades/Blade Center Again it is a matter of intra‐blade center networking bandwidth and latency issues in the link connecting the blades, like the backplane……….Carnegie MellonSpring 2010 ©15-319 Introduction to Cloud Computing# Processors/Blade  Each Processor has cores, L1 & L2 caches. Both caches connected to/share/retrieve/write data to same L3 cache.  The network pipe between processors & L3 cache requires bandwidth and latency good enough to serve all processors. Also, A delay comes with needing a connection/internal bus between processors Especially if they were to do parallel processing and exchange data. Having the data on the same processor chip would have saved this communication & datatransformation delay.L1L2L1L2L3RAMDiskCORESCORESSimple view of a dual-core bladeCarnegie MellonSpring 2010 ©15-319 Introduction to Cloud Computing# Cores/Processor Each processor has cores that share L1 cache & L2 cache.  They are shared resources between the cores.  Again bandwidth & latency of the network pipes connecting the cores to the caches affects the overall functionality.  There are single‐core, dual‐core, quad‐core , …. processors  A blade with multiple processors is called Symmetric Multi‐ processor Machine (SMP)L1L2L1L2L3RAMDiskCORESCORESSimple view of a dual-core bladeCarnegie MellonSpring 2010 ©15-319 Introduction to Cloud ComputingWhy a Multi‐core Processor? Problem: How much faster can a processor get? Towards the early 2000’s we hit a “frequency wall” for processors. Processor frequencies topped out at 4 GHz due to the thermal limit of Silicon.© Christian ThostrupCarnegie MellonSpring 2010 ©15-319 Introduction to Cloud ComputingWhy a Multi‐core Processor? Solution : More chips in Parallel! Enter MulticoreMulticore.. Cram more processing units into the chip rather than increase Cram more processing units into the chip rather than increase clock frequency.clock frequency. Traditional processors are available with up to Quad and Six CorTraditional processors are available with up to Quad and Six Cores.es. The Cell Broadband Engine (The Cell Broadband Engine (PlaystationPlaystation 3) has 9 cores.3) has 9 cores. Intel recently announced a 48Intel recently announced a 48‐‐core processor.core processor.©IntelCarnegie MellonSpring 2010 ©15-319 Introduction to Cloud ComputingMulticore processors Have fully functioning processor “cores” in a processor. Have individual L1 and shared L2 caches.  OS and applications see each core as an individual processor. Applications have to be specifically rewritten for optimized performance.Source: hardwaresecrets.comCarnegie MellonSpring 2010 ©15-319 Introduction to Cloud ComputingDesign Concerns: processor frequency Frequency It is the clock speed of the processor  The faster the clock speed, the more instruction the processor can execute at a given timeCarnegie MellonSpring 2010 ©15-319 Introduction to Cloud ComputingCore Design (1/4) Order of superscalars Superscalar of a core is divided at levels: A level‐x core is the one with x functional units.  The more functional units: the more parallelism the system can do & the less is the time