Token Coherence

1
Token Coherence

• Milo M. K. Martin, Mark D. Hill, and David A.
  Wood
• Presented By
• Jerry Wu

2
Introduction

• Goals
• Effectively utilize glueless design for
  multiprocessor servers
• Implement a low-latency cache coherence protocol
  that is also correct
• Separating performance from correctness

3
Problems and Solution

• Traditional Snooping
• Requires a totally-ordered interconnect
• Directory
• Indirection latency for cache-to-cache misses
• Ideal coherence protocol
• Avoid both indirections and interconnect ordering
• This approach suffers from numerous race cases
  that are difficult to make correct
• Solution Token Coherence!

4
Token Coherence Architecture

• Two halves to the architecture
• Correctness substrate
• Uses token counting to enforce safety and uses
  persistent requests to prevent starvation
• Allows the movement of data around the system
  without concern for order or races
• Performance protocol
• Uses hint requests to direct the correctness
  substrate to send data and tokens to the
  requesting processor
• TokenB protocol Token-Coherence-using-Broadcast

5
Correctness Substrate

• Enforcing Safety via Token Counting
• Each block has T tokens at all time, one of which
  is the owner token
• MOSI protocol
• Processor can write only if it has all T tokens
  (M)
• Processor can read only if it has at least one
  token (S)
• Processor has no tokens (I)
• Processor has the owner token but not all other
  tokens (O), if message contains the owner token,
  it must contain data
• Valid bit is used to distinguish non-owner tokens
  without valid data

6
Correctness Substrate II

• Avoiding Starvation via Persistent Requests
• Initiated when possible starvation is detected
• At most one persistent request activated at any
  time
• All nodes sees the persistent request and forward
  all tokens for the block
• Memory operation is performed by the initiator
  before deactivating the persistent request

7
Performance Protocol

• No obligations for correctness! Make the common
  case fast (Amdahls Law).
• Transient requests fast, unordered hint
  requests which may fail
• TokenB protocol
• Processors broadcast all transient requests
• Respond to transient requests using common MOSI
  protocol
• Transient requests are reissued when failed until
  persistent request is activated

8
Evaluation interconnection network

• Two different interconnection networks are used
  in evaluating token coherence vs. snooping.
• 2-level tree is used for totally-ordered
  interconnection
• 2D torus is used for unordered interconnect

9
Evaluation

• Is the number of reissued/persistent requests
  small?
• Yes. 97 of cache misses are issued only once
• Can TokenB outperform Snooping?
• Yes. TokenB on Torus is faster than Snooping on
  Tree by 15-28
• Can TokenB outperform Directory and Hammer?
• Yes. Removing indirection from critical path
  makes TokenB 17-54 faster than Directory and
  8-29 faster than Hammer
• TokenBs traffic compared with Directory and
  Hammer?
• Hammer uses 79-90 more traffic, Directory uses
  21-25 less traffic
• Is TokenB protocol scalable?
• No. Broadcasting limits its scalability.

10
Questions?