Processor Architecture: Current Trends A B Transfer a truckload at - - PowerPoint PPT Presentation

Processor Architecture: Current Trends A B Transfer a truckload at a time from A to B

Processor Architecture: Current Trends A B Make it more powerful Faster: Engine, Lighter, Wheels, ….. Bigger: More work per trip

Processor Architecture: Current Trends A B

Make it more powerful Faster: Engine, Lighter, Wheels, …. Bigger: More work per trip.

Evolutionary improvements in technology reach their limit Frequency Power Pipeline depth Insufficient work

Processor Architecture: Current Trends

• Instruction Level Parallelism (ILP)
• Concurrency between instructions in a single program
• Application is oblivious to underlying parallelism
• Hardware mechanisms and compiler optimizations
• Simple Instruction Pipelining
• Advanced Pipelining Techniques
• Superscalar and VLIW architectures
• Inflection Point: Current ILP-based designs have reached a plateau
• Power Wall
• Memory Wall
• Implicit Parallelism Wall

Processor Architecture: Current Trends A B Get a fleet of trucks Increase Throughput Rate of package delivery

Processor Architecture: Current Trends A B Multiple cores on a chip Cores are simple Great if coordination between cores is small

Get a fleet of trucks

Processor Architecture: Current Trends

• Explicit Parallelism

Software ( application, compiler, run-time) aware of parallelism

• Data Level Parallelism (DLP)
• Identical operations on elements of large arrays (vectors)
• Vector or Array Processors
• GPUs
• Thread Level Parallelism (TLP)
• Applications with multiple cooperating or independent threads
• Multithreaded Processors
• Chip Multiprocessors
• Process Level Parallelism
• Scalable Multiprocessors, Clusters (Established technology)
• Application Level Parallelism
• Virtualization, Cloud Computing?

Processor Architecture: Current Trends A B Move a single package from A to B Having multiple trucks does not help Task cannot be obviously parallelized High Performance Scalar Processor Interested in reducing the time for a non-decomposable task Reduce Latency

Memory Subsystem

• Memory Hierarchy is key to performance
• Latency Hiding
• Caching principles
• Cache organization
• Multi-level caching
• Virtual Memory
• Interaction of Caching and Virtual Memory
• Multiprocessor Cache Coherence
• Memory Consistency

I/O Subsystem

• Storage : Evolution from Systems to Service
• Direct or Server-Attached Storage:
• Storage devices : Disk Drives, SSDs, MEMS …
• Interconnect protocols (IDE, SATA, SCSI, ..)
• Networked Storage
• Storage Arrays
• Array Virtualization
• Storage Area Networks (SAN), Networked Attached Storage (NAS)
• Protocols: Fiber Channel, iSCSI, RDMA
• Fault Tolerance:
• RAID Arrays, Replication, Erasure Codes for Multiple errors
• Storage as a Service: Amazon S3, Google G-Drive, Apple, …
• Isolation and Performance Guarantees
• Interconnect
• Bus systems
• Point-to-point interconnect
• Switched fabrics
• Cluster interconnects