Episode 44: Interview Brian Goetz and David Holmes

Data Structures

Concurrent data structures are designed to be thread-safe and highly performant, but they come with their own set of challenges. David Holmes explains that while these structures offer high performance, they may lack the exclusivity needed in certain applications 1. He notes that using non-blocking algorithms, which rely on atomic instructions instead of locks, can simplify data structures but are extremely difficult to implement correctly 1. Brian Goetz adds that the complexity of developing lock-free algorithms makes them suitable only for experts, and their benefits are best realized when encapsulated in system libraries 1.

If writing correct programs with locks is hard, writing non-blocking algorithms is ten times harder.

--- Brian Goetz

Robert Blumen suggests partitioning data structures to avoid locking entire structures, such as a hash table, which can improve performance 2.

   

Transactional Memory

Transactional memory offers a promising approach to concurrency control by allowing programmers to specify desired outcomes rather than implementation details. Brian Goetz describes it as a model still in the research phase, aiming to simplify concurrency by letting the system determine the most efficient execution method 3. This approach contrasts with traditional locking mechanisms, which often complicate algorithms and increase the risk of deadlock 4.

The idea is to allow the programmer to specify what he wants, not how he wants to do it.

--- Brian Goetz

By eliminating the need for programmers to define the scope of data contention, transactional memory could potentially offer scalability and simplicity similar to garbage collection in programming 4.