ManyCore Architectures

1
Many-Core Architectures

• Keith Lowery
• AMD

2
Workload Characteristics

• Today relatively coarse-grained, long-lived
  parallel threads.
• Tomorrow - Fine-grained short-lived parallel
  units that are much lighter-weight than todays
  threads.
• Burstiness
• Adaptive exploitation of available cores for
  parallelization.
• Wider embrace of non-blocking approaches to
  synchronization.

3
Scheduling

• Scheduling in the context of a surplus of cores
  is not the scheduling of today
• Fairness will be an unnecessary (and
  inefficient) concern for most of the available
  cores.
• Run-to-completion will be more appropriate for
  many types of workloads.
• Parallelization patterns employed by
  highly-parallel applications will entail
  scheduling behaviors that take application-specifi
  c patterns into account.
• Will be very difficult to maximize cache hit
  rates and reduce interconnect saturation without
  some understanding of the parallelization
  pattern(s) being employed.
• Probably implies that user-mode scheduling is
  more appropriate.

4
Multi-Threading is for Servers Only?

• Synergistic relationship between size of data and
  benefits of parallelization
• Increasing availability of economical bandwidth
  and storage bodes well for many-core
  opportunities on the client.
• Intuition underestimates the size of data that
  clients are already working with which can be
  tapped to introduce qualitative changes to the
  user experience.
• Tension between personalization and privacy
  applies pressure to move personalization
  analytics to the client.
• Attention needs to be paid to the extent to which
  networking protocols either encourage or obstruct
  parallelization.

5
Programming Model for Many-Core

• Hey Hey, Ho Ho, locking semantics have got to go.
• May be the single biggest obstacle to broader
  utilization of parallelism in business
  applications.
• Significant businesses have declared
  multi-threaded programming to be evil and
  forbidden development as a result.
• Non-blocking approaches to synchronization can
  add anywhere from .5x to 4X latency to
  synchronized memory access in applications.
• Ripe for hardware assistance.

6
Programming Model for Many-Core

• Just-in-Case Execution
• Application-level transactional semantics may
  combine well with continuous analytics and a
  surplus of cores to do speculative execution.
• Must be apprehendable by developers who are
  primarily domain experts.