High Performance Computing 25 SP Programming SMP Platform

Sharing address space brings simplification.

Shared Address Space Programming Paradigm

Vary on mechanism for data sharing, concurrency models and support for synchronization.

• Process based model
• Lightweight processes and threads.

Thread

A thread is a single stream of control in the flow of a program.

Logical memory model of a thread:

All memory is globally accessible to every thread and threads are invoked as function calls.

Benefits of threads:

• Take less time to create a new thread than a new process.
• Less time to terminate a thread than a process.

The taxonomy of thread:

• User-Level Thread(ULT): The kernel is not aware of the existence of threads. All thread management is done by the application using a thread library. Thread switching does not require kernel mode privileges and scheduling is application specific.
• Kernel-Level Thread(KLT): All thread management is done by kernel. No thread library but an API to the kernel thread facility. Switching between threads requires the kernel and schedule on a thread basis.
• Combined ULT and KLT approaches. Combines the best of both approaches.

PThread Programming

Potential problems with threads:

• Conflicting access to shared memory.
• Race condition occur.
• Starvation
• Priority inversion
• Deadlock

Mutual Exclusion

Mutex locks: implementing critical sections and atomic operations.

Two states: locked and unlocked. At any point of time, only one thread can lock a mutex lock.

Producer-Consumer Work Queues

The producer creates tasks and inserts them into a work-queue.

The consumer threads pick up tasks from the task queue and execute them.

Locks represent serialization points since critical sections must be executed by threads one after the other.

Important: Minimize the size of critical sections.

Condition Variables for Synchronization

The pthread_mutex_trylock alleviates the idling time but introduce the overhead of polling for availability of locks.

An interrupt driven mechanism as opposed to a polled mechanism as the availability is signaled.

A condition variable: a data object used for synchronizing threads. Block itself until specified data reaches a predefined state.

When a thread performs a condition wait, it's not runnable as not use any CPU cycles but a mutex lock consumes CPU cycles as it polls for the locks.

Common Errors: One cannot assume any order of execution, must be explicitly established by mutex, condition variables and joins.

MPI Programming

Low cost message passing architecture.

Mapping of MPI Processes:

MPI views the processes as a one-dimensional topology. But in parallel programs, processes are arranged in higher-dimensional topologies. So it is required to map each MPI process to a process in the higher dimensional topology.

Non-blocking Send and Receive:

MPI_ISend and MPI_Irecv functions allocate a request object and return a pointer to it.

OpenMP

A standard for directive based parallel programming.

Thread based parallelism and explicit parallelism.

Use fork-join model: