Concurrent and Distributed Systems

Aims

This course considers two closely related topics, Concurrent Systems and Distributed Systems, over 16 lectures. The aim of the first half of the course is to introduce concurrency control concepts and their implications for system design and implementation. The aims of the latter half of the course are to study the fundamental characteristics of distributed systems, including their models and architectures; the implications for software design; some of the techniques that have been used to build them; and the resulting details of good distributed algorithms and applications.

Lectures: Concurrency

• Introduction to concurrency, threads, and mutual exclusion. Introduction to concurrent systems; threads; interleaving; preemption; parallelism; execution orderings; processes and threads; kernel vs. user threads; M:N threads; atomicity; mutual exclusion; and mutual exclusion locks (mutexes).
• Automata Composition. Synchronous and asynchronous parallelism; sequential consistency; rendezvous. Safety, liveness and deadlock; the Dining Philosophers; Hardware foundations for atomicity: test-and-set, load-linked/store-conditional and fence instructions. Lamport bakery algorithm.
• Common design patterns: semaphores, producer-consumer, and MRSW. Locks and invariants; semaphores; condition synchronisation; N-resource allocation; two-party and generalised producer-consumer; Multi-Reader, Single-Write (MRSW) locks.
• CCR, monitors, and concurrency in practice. Conditional critical regions (CCR); monitors; condition variables; signal-wait vs. signal-continue semantics; concurrency in practice (kernels, pthreads, Java).
• Deadlock and liveness guarantees Offline vs. online; model checking; resource allocation graphs; lock order checking; deadlock prevention, avoidance, detection, and recovery; livelock; priority inversion; priority inheritance.
• Concurrency without shared data; transactions. Active objects; message passing; tuple spaces; CSP; and actor models. Composite operations; transactions; ACID; isolation; and serialisability.
• Further transactions History graphs; good and bad schedules; isolation vs. strict isolation; 2-phase locking; rollback; timestamp ordering (TSO); and optimistic concurrency control (OCC).
• Crash recovery, lock-free programming, and transactional memory. Write-ahead logging, checkpoints, and recovery. Lock-free programming. Hardware and software transactional memories.

Lectures: Distributed Systems

• Introduction to distributed systems; RPC. Avantages and challenges of distributed systems; unbounded delay and partial failure; network protocols; transparency; client-server systems; remote procedure call (RPC); marshalling; interface definition languages (IDLs).
• System models and faults. Synchronous, partially synchronous, and asynchronous network models; crash-stop, crash-recovery, and Byzantine faults; failures, faults, and fault tolerance; two generals problem.
• Time, clocks, and ordering of events. Physical clocks; UTC; clock synchronisation, drift, and compensation; Network Time Protocol (NTP). Logical time; happens-before relation; Lamport clocks; vector clocks.
• Replication. Fault tolerance; leader-based, multi-leader, and leaderless replication; quorum systems; replica consistency; linearizability; CAP theorem; eventual consistency; session guarantees.
• Middleware and protocols. Process groups; FIFO, causal order, and total order broadcast; message-oriented and object-oriented middleware; distributed mutual exclusion.
• Consensus and distributed transactions. Leader elections; consensus; the FLP result; Paxos and Raft; state machine replication; distributed transactions; atomic commit protocols; 2-phase commit.
• Case studies. Network File System (NFS); Amazon’s Dynamo; Google datacentre technologies (e.g. MapReduce, Spanner); cloud computing services.
• Advanced topics. Conflict-free Replicated Data Types (CRDTs); Byzantine fault tolerance; peer-to-peer systems; distributed systems security.

Objectives

At the end of Concurrent Systems portion of the course, students should:

• understand the need for concurrency control in operating systems and applications, both mutual exclusion and condition synchronisation;
• understand how multi-threading can be supported and the implications of different approaches;
• be familiar with the support offered by various programming languages for concurrency control and be able to judge the scope, performance implications and possible applications of the various approaches;
• be aware that dynamic resource allocation can lead to deadlock;
• understand the concept of transaction; the properties of transactions, how they can be implemented, and how their performance can be optimised based on optimistic assumptions;
• understand how the persistence properties of transactions are addressed through logging; and
• have a high-level understanding of the evolution of software use of concurrency in the operating-system kernel case study.

At the end of the Distributed Systems portion of the course, students should:

• understand the difference between shared-memory concurrency and distributed systems;
• understand the fundamental properties of distributed systems and their implications for system design;
• understand notions of time, including logical clocks, vector clocks, and physical time synchronisation;
• be familiar with various approaches to data and service replication, as well as the concept of data consistency;
• understand the effects of large scale on the provision of fundamental services and the tradeoffs arising from scale;
• appreciate the implications of individual node and network communications failures on distributed computation;
• be aware of a variety of programming models and abstractions for distributed systems, such as RPC, middleware, and total order broadcast;
• be familiar with a range of distributed algorithms, such as consensus and two-phase commit;
• be familiar with a number of case studies in distributed-system design including the Network File System (NFS), the Network Time Protocol (NTP), Amazon's Dynamo, and Google’s MapReduce and Spanner systems.

Recommended reading

• Bacon, J. and Harris, T. (2003). Operating systems: distributed and concurrent software design. Addison-Wesley.
• Bacon, J. (1997). Concurrent systems. Addison-Wesley.
• Kleppmann, M. (2017). Designing data-intensive applications. O’Reilly.
• Tanenbaum, A.S. and van Steen, M. (2017). Distributed systems, 3rd edition. available online.
• Cachin, C., Guerraoui, R. and Rodrigues, L. (2011) Introduction to Reliable and Secure Distributed Programming. Springer (2nd edition).