Operating System Support for Fine-Grain Parallelism on Multicore Architectures

1
Operating System Support forFine-Grain
Parallelism on Multicore Architectures
John Giacomoni

• Manish Vachharajani
• University of Colorado at Boulder
• 2007.10.14

2
Problem

• UP performance at end of life
• Chip-Multiprocessor systems
• What do we want from multicore systems?

• Individual cores less powerful than UP
• Asymmetric and Heterogeneous
• 10s-100s-1000s of cores

Performance!
Intel (2x2-core)
MIT RAW (16-core)
100-core
400-core
3
ExtractingPerformance

• Task Parallelism
• Desktop
• Data Parallelism
• Web serving
• Split/Join, MapReduce, etc
• Pipeline Parallelism
• Video decoding
• Network processing

4
ExtractingPerformance (2)

• Stream Parallelism
• Combines
• Data Parallelism
• Pipeline Parallelism
• Ad-Hoc Parallelism
• Semi- or unstructured
• Usual thread model

5
Focus onPipeline Parallelism

• Most stringent timing requirements
• Example applications
• Network Processing
• Network Intrusion Detection
• DDoS Filtering
• Multimedia processing
• Transcoding
• Signal Processing
• Software Defined Radio
• Also applies to
• Data parallelism
• Stream Parallelism

6
Soft Network Processing(Soft-NP)

• How do we protect?

• GigE Network Properties
• 1,488,095 frames/sec
• 672 ns/frame
• Frame dependencies

Frame Shared Memory Line-Rate Networking on
Commodity Hardware. To Appear Proceedings of
the ACM/IEEE Symposium on Architectures for
Networking and Communications Systems 2007
(ANCS), December 2007. John Giacomoni, John K.
Bennett, Antonio Carzaniga, Douglas C. Sicker,
Manish Vachharajani and Alexander L. Wolf.
7
Frame Shared Memory(Soft-NP)
Input (IP) Output(OP)
8
What OS support is necessary?
9
Low-OverheadCommunication
Gigabit Ethernet
Syscalls 170ns
pthread mutex 200ns
10
FastForward

• Portable software only framework
• 35-40ns/queue operation 2.0 GHz AMD Opteron
• 26-28ns/queue operation 2.6 GHz AMD Opteron
• Architecturally tuned CLF queues
• Works with strong to weak consistency models
• Hides die-die communication
• Robust against unbalanced stages
• Poster FastForward for Efficient Pipeline
  Parallelism. Proceedings of the 16th
  International Conference on Parallel
  Architectures and Compilation Techniques (PACT),
  September 2007. John Giacomoni, Tipp Moseley,
  Manish Vachharajani.

11
FastForwardPerformance
Lamport
FF
FF Unbalanced
FF Re-Balanced
12
Zero-StallGuarantee
13
GangScheduling

• Optimize for application performance
• Instead of system throughput or fairness
• Computer Utility -gt max(System Utilization)
• Multicore system -gt excess of resources.
• Dedicate resources to pipeline applications
• Want selective timesharing

14
SystemServices

• Fast!
• Synchronous calls introduce too much overhead
• System calls 170ns
• Asynchronous calls may limit parallelism
• Want System services with independent I/O paths

15
PipelinableSystem Services

• Mixing stages from multiple process domains
• Push model vs. call/return or poll
• Hardware can be an active participant

16
HeterogeneousGang Scheduling

• Need a single scheduling label for every pipeline
  stage
• Ensures simultaneous scheduling of every
  necessary resource
• (zero-stall guarantee)
• Including hardware stages.
• Scheduling multi-domain entities

17
Multi-DomainEntities

• Application state
• Shared with local stages
• Pipeline private state
• Stage state shared with pipeline and parent
  process.
• The multi-domain application model respects the
  private data model implicit in single-domain
  applications while providing first-class naming
  for multi-domain pipelines.

18
Summaryof Discussion

1. Low-overhead communication
2. Zero-stall guarantee
3. Selective timesharing
4. Pipelineable system services
5. Heterogenous gang scheduling
6. Pipelines as multi-domain applications

19
Questions?
john.giacomoni_at_colorado.edu