Memory Coherence in Shared Virtual Memory Systems

1
Memory Coherence in Shared Virtual Memory Systems

• Kai Li and Paul Hudak
• Presentation G. Passas

2
Problem and Design Choices

• Coherence granularity 1page (2KB)
• Page synchronization invalidation (instead of
  writeback)
• Page ownership dynamic (instead of static)
• Coherence violations cause page faults and
• Page fault handlers implement the coherence
  protocols.

3
A centralized protocol for coherent SVM
Page entry at local page table
Page entry at manager info table
4
Distributing the managerial task

• Move synchronization of page ownership totally to
  the individual owners

confirmation message is eliminated

• Keep track of ownership at each processors
  local page table

• Map pages to processors in a fixed manner
• Each processor knows a probable owner

5
A distributed protocol for coherent SVM
Page entry at local page table

• Resembles directories
• At most N-1 messages for each req

6
The shortcomings of SVM

• Great false sharing due to coherence granularity
• frequent page transfers of large units
  between nodes
• The interrupt cost for a message reception is high

7
Implementation and performance of Munin

• J. Karpet, J. Bennet and W. Zwaenepoel

Presentation G. Passas
8
Core Ideas

• Release consistency to amortize the cost of
  message transfers

Propagate invalidations/updates only after a
release

• Multiple consistency (coherence)
• protocols

Cope with different access paterns
Annotate shared variables to indicate preferred
protocol
The core system is basically the same with the
distributed system previously described
9
Annotations and protocol params
10
Implementation issues

• How are page tables (object directories)
  generated?
• (preprocessor, auxiliary files, linking, root
  Munin thread (fault handler)
• How are write updates/invalidations delayed?
• (Delayed update queue, DUQ. Flushed upon a
  lock release)
• How is synchronization implemented?
• (queue based locks, synch. object directory)

11
Performance vs. message passing

• Benchmark multiplication of 2 400x400 matrices
• User time time executing user core
• System time time executing Munin code

12
Shortcomings

• Constraints to the programmer
• (A) Annotations
• (B) Use of the systems synchronization libs