Explore the concept of time traveling hardware and software systems in this Stanford seminar. Delve into technology scaling, various types of parallelism, and the challenges posed by dependencies. Examine the notion of correctness and sequential consistency in computing systems. Investigate shared memory systems, directory-based coherence, and cache coherence scalability. Learn about lease-based coherence and logical timestamps through practical examples. Analyze the pros and cons of the TARDIS system and explore concurrency control techniques. Discover innovative approaches to timestamp management and transaction execution in the TicToc system. Gain insights into physiological time across the computing stack and its implications for future hardware and software design.
Overview
Syllabus
Introduction.
TECHNOLOGY SCALING.
DIFFERENT KINDS OF PARALLELISM - 1.
DIFFERENT KINDS OF PARALLELISM - 2.
DEPENDENCY DESTROYS PARALLELISM.
DIFFERENT KINDS OF DEPENDENCY.
DEPENDENCE IS ACROSS TIME, BUT WHAT IS TIME?.
WAR DEPENDENCE Initially A = 10.
WHAT IS CORRECTNESS?.
SEQUENTIAL CONSISTENCY.
AVOIDING DEPENDENCY ACROSS THE STACK.
SHARED MEMORY SYSTEMS.
DIRECTORY-BASED COHERENCE.
CACHE COHERENCE SCALABILITY.
LEASE-BASED COHERENCE.
LOGICAL TIMESTAMP.
TWO-CORE EXAMPLE.
STORE A @ CORE O.
LOAD B @ CORE O.
STORE B @ CORE 1.
TWO VERSIONS COEXIST.
LOAD A @ CORE 1.
SUMMARY OF EXAMPLE Directory.
TARDIS PROS AND CONS.
CONCURRENCY CONTROL.
BOTTLENECK 1: TIMESTAMP ALLOCATION.
BOTTLENECK 2: STATIC ASSIGNMENT.
KEY IDEA: DATA DRIVEN TIMESTAMP MANAGEMENT.
TicToc TRANSACTION EXECUTION.
LOAD A FROM T1 2.
COMMIT PHASE OF T1.
FINAL STATE.
PHYSIOLOGICAL TIME ACROSS THE STACK.
Taught by
Stanford Online
