Parallel Computing: Theory and Practice

We assume a machine model that consists of a shared memory by a number of processors, usually written as $P$. The processors have access to a shared memory, which is readable and writable by all processors.

We assume that an operating system or a that allows us to create processes. The kernel schedules processes on the available processors in a way that is mostly out of our control with one exception: the kernel allows us to create any number of processes and pin them on the available processors as long as no more than one process is pinned on a processor.

We define a thread to be a piece of sequential computation whose boundaries, i.e., its start and end points, are defined on a case by case basis, usually based on the programming model. In reality, there different notions of threads. For example, a system-level thread is created by a call to the kernel and scheduled by the kernel much like a process. A user-level thread is created by the application program and is scheduled by the application—user level threads are invisible to the kernel. Common property of all threads is that they perform a sequential computation. In this class, we will usually talk about user-level threads. In the literature, you will encounter many different terms for a user-level thread, such as "fiber", "sparc", "strand", etc.

For our purposes in this book an application, a piece of runnable software, can only create threads but no processes. We will assume that we can assign to an application any number of processes to be used for execution. If an application is run all by itself (without any other application running at the same time) and if all of its processes are pinned, then we refer to such an execution as occurring in the dedicated mode.

This book is entirely based on C++ and a library for writing parallel programs in C++. We use recent features of C++ such as closures or lambda expressions and templates. A deep understanding of these topics is not necessary to follow the course notes, because we explain them at a high level as we go, but such prior knowledge might be helpful; some pointers are provided below.

A multithreaded computation can be represented by a dag, a Directed Acyclic Graph, or written also more simply a dag, of vertices. The figure below show an example multithreaded computation and its dag. Each vertex represents the execution of an instruction, such as an addition, a multiplication, a memory operation, a (thread) spawn operation, or a synchronization operation. A vertex representing a spawn operation has outdegree two. A synchronization operation waits for an operation belonging to a thread to complete, and thus a vertex representing a synchronization operation has indegree two. Recall that a dag represents a partial order. Thus the dag of the computation represents the partial ordering of the dependencies between the instructions in the computation.

Throughout this book, we make two assumptions about the structure of the dag:

The outdegree assumption naturally follows by the fact that each vertex represents an instruction, which can create at most one thread.

For analyzing the efficiency and performance of multithreaded programs, we use several cost measures, the most important ones include work and span. We define the work of a computation as the number of vertices in the dag and the span as the length of the longest path in the dag. In the example dag above, work is $15$ and span is $9$.

The execution of a multithreaded computation executes the vertices in the dag of the computation in some partial order that is consistent with the partial order specified by the dag, that is, if vertices $u,v$ are ordered then the execution orders them in the same way.

Multithreaded programs are executed by using a scheduler that assigns vertices of the dag to processes.

The first condition ensures that a schedule observes the dependencies in the dag. Specifically, for each arc $(u,v)$ in the dag, the vertex $u$ is executed before vertex $v$.

For any step in the execution, we call a vertex ready if all the ancestors of the vertex in the dag are executed prior to that step. Similarly, we say that a thread is ready if it contains a ready vertex. Note that a thread can contain only one ready vertex at any time.

The first lower bound follows by the simple observation that a schedule can only execute $P$ instructions at a time. Since all vertices must be executed, a schedule has length at least $\frac{W}{P}$. The second lower bound follows by the observation that the schedule cannot execute a vertex before its ancestors and thus the length of the schedule is at least as long as any path in the dag, which can be as large as the span $S$.

Having established a lower bound, we now move on to establish an upper bound for the offline scheduling problem, where we are given a dag and wish to find an execution schedule that minimizes the run time. It is known that the related decision problem in NP-complete but that 2-approximation is relatively easy. We shall consider two distinct schedulers: level-by-level scheduler and greedy scheduler.

A level-by-level schedule is a schedule that executes the instructions in a given dag level order, where the level of a vertex is the longest distance from the root of the dag to the vertex. More specifically, the vertices in level 0 are executed first, followed by the vertices in level 1 and so on.

This theorem called Brent’s theorem was proved by Brent in 1974. It shows that the lower bound can be approximated within a factor of $2$.

Brent’s theorem has later been generalized to all greedy schedules. A greedy schedule is a schedule that never leaves a process idle unless there are no ready vertices. In other words, greedy schedules keep processes as busy as possibly by greedily assigning ready vertices.

In offline scheduling, we are given a dag and are interested in finding a schedule with minimal length. When executing multithreaded program, however, we don’t have full knowledge of the dag. Instead, the dag unfolds as we run the program. Furthermore, we are interested in not minimizing the length of the schedule but also the work and time it takes to compute the schedule. These two additional conditions define the online scheduling problem.

An online scheduler or a simply a scheduler is an algorithm that solves the online scheduling problem by mapping threads to available processes. For example, if only one processor is available, a scheduler can map all threads to that one processor. If two processors are available, then the scheduler can divide the threads between the two processors as evenly as possible in an attempt to keep the two processors as busy as possible by load balancing.

There are many different online-scheduling algorithms but these algorithms all operate similarly. We can outline a typical scheduling algorithm as follows.

For a given schedule generated by an online scheduling algorithm, we can define a tree of vertices, which tell us far a vertex, the vertex that enabled it.

The centralized scheduler with the global thread queue is a greedy scheduler that generates a greedy schedule under the assumption that the queue operations take zero time and that the dag is given. This algorithm, however, does not work well for online scheduling the operations on the queue take time. In fact, since the thread queue is global, the algorithm can only insert and remove one thread at a time. For this reason, centralized schedulers do not scale beyond a handful of processors.

There has been much research on the problem of reducing friction in scheduling. This research shows that distrubuted scheduling algorithms can work quite well. In a distributed algorithm, each processor has its own queue and primarily operates on its own queue. A load-balancing technique is then used to balance the load among the existing processors by redistributing threads, usually on a needs basis. This strategy ensures that processors can operate in parallel to obtain work from their queues.

A specific kind of distributed scheduling technique that can leads to schedules that are close to optimal is work stealing schedulers. In a work-stealing scheduler, processors work on their own queues as long as their is work in them, and if not, go "steal" work from other processors by removing the thread at the tail end of the queue. It has been proven that randomized work-stealing algorithm, where idle processors randomly select processors to steal from, deliver close to optimal schedules in expectation (in fact with high probability) and furthermore incur minimal friction. Randomized schedulers can also be implemented efficiently in practice. PASL uses an scheduling algorithm that is based on work stealing. We consider work-stealing in greater detail in a future chapter.

Multithreaded programs can be written using a variety of language abstractions interfaces. One of the most widely used interfaces is the POSIX Threads* or *Pthreads interface, which specifies a programming interface for a standardized C language in the IEEE POSIX 1003.1c standard. Pthreads provide a rich interface that enable the programmer to create multiple threads of control that can synchronize by using the nearly the whole range of the synchronization facilities mentioned above.

An example Pthread program is shown below. The main thread (executing function main) creates 8 child threads and terminates. Each child in turn runs the function helloWorld and immediately terminates. Since the main thread does not wait for the children to terminate, it may terminate before the children does, depending on how threads are scheduled on the available processors.

When executed this program may print the following.

But that would be unlikely, a more likely output would look like this:

And may even look like this

Interface such as Pthreads enable the programmer to create a wide variety of multithreaded computations that can be structured in many different ways. Large classes of interesting multithreaded computations, however, can be expcessed using a more structured approach, where threads are restricted in the way that they synchronize with other threads. One such interesting class of computations is fork-join computations where a thread can spawn or "fork" another thread or "join" with another existing thread. Joining a thread is the only mechanism through which threads synchronize. The figure below illustrates a fork-join computation. The main threads forks thread A, which then spaws thread B. Thread B then joins thread A, which then joins Thread M.

In addition to fork-join, there are other interfaces for structured multithreading such as async-finish, and futures. These interfaces are adopted in many programming languages: the Cilk language is primarily based on fork-join but also has some limited support for async-finish; X10 language is primarily based on async-finish but also supports futures; the Haskell language Haskell language provides support for fork-join and futures as well as others; Parallel ML language as implemented by the Manticore project is primarily based on fork-join parallelism. Such languages are sometimes called implicitly parallel.

The class computations that can be expressed as fork-join and async-finish programs are sometimes called nested parallel. The term "nested" refers to the fact that a parallel computation can be nested within another parallel computation. This is as opposed to flat parallelism where a parallel computation can only perform sequential computations in parallel. Flat parallelism used to be common technique in the past but becoming increasingly less prominent.

Structured multithreading offers important benefits both in terms of efficiency and expressiveness. Using programming constructs such as fork-join and futures, it is usually possible to write parallel programs such that the program accepts a "sequential semantics" but executes in parallel. The sequential semantics enables the programmer to treat the program as a serial program for the purposes of correctness. A run-time system then creates threads as necessary to execute the program in parallel. This approach offers is some ways the best of both worlds: the programmer can reason about correctness sequentially but the program executes in parallel. The benefit of structured multithreading in terms of efficiency stems from the fact that threads are restricted in the way that they communicate. This makes it possible to implement an efficient run-time system.

More precisely, consider some sequential language such as the untyped (pure) lambda calculus and its sequential dynamic semantics specified as a strict, small step transition relation. We can extend this language with the structured multithreading by enriching the syntax language with "fork-join" and "futures" constructs. We can now extend the dynamic semantics of the language in two ways: 1) trivially ignore these constructs and execute serially as usual, and 2) execute in parallel by creating parallel threads. We can then show that these two semantics are in fact identical, i.e., that they produce the same value for the same expressions. In other words, we can extend a rich programming language with fork-join and futures and still give the language a sequential semantics. This shows that structured multithreading is nothing but an efficiency and performance concern; it can be ignored from the perspective of correctness.

We use the term parallelism to refer to the idea of computing in parallel by using such structured multithreading constructs. As we shall see, we can write parallel algorithms for many interesting problems. While parallel algorithms or applications constitute a large class, they don’t cover all applications. Specifically applications that can be expressed by using richer forms of multithreading such as the one offered by Pthreads do not always accept a sequential semantics. In such concurrent applications, threads can communicate and coordinate in complex ways to accomplish the intended result. A classic concurrency example is the "producer-consumer problem", where a consumer and a producer thread coordinate by using a fixed size buffer of items. The producer fills the buffer with items and the consumer removes items from the buffer and they coordinate to make sure that the buffer is never filled more than it can take. We can use operating-system level processes instead of threads to implement similar concurrent applications.

In summary, parallelism is a property of the hardware or the software platform where the computation takes place, whereas concurrency is a property of the application. Pure parallelism can be ignored for the purposes of correctness; concurrency cannot be ignored for understanding the behavior of the program.

Parallelism and concurrency are orthogonal dimensions in the space of all applications. Some applications are concurrent, some are not. Many concurrent applications can benefit from parallelism. For example, a browser, which is a concurrent application itself as it may use a parallel algorithm to perform certain tasks. On the other hand, there is often no need to add concurrency to a parallel application, because this unnecessarily complicates software. It can, however, lead to improvements in efficiency.

The following quote from Dijkstra suggest pursuing the approach of making parallelism just a matter of execution (not one of semantics), which is the goal of the much of the work on the development of programming languages today. Note that in this particular quote, Dijkstra does not mention that parallel algorithm design requires thinking carefully about parallelism, which is one aspect where parallel and serial computations differ.

Now, we have all the tools we need to describe our first parallel code: the recursive Fibonacci function. Although useless as a program because of efficiency issues, this example is the "hello world" program of parallel computing.

Recall that the $n^{th}$ Fibonnacci number is defined by the recurrence relation

with base cases

Let us start by considering a sequential algorithm. Following the definition of Fibonacci numbers, we can write the code for (inefficiently) computing the $n^{th}$ Fibonnacci number as follows. This function for computing the Fibonacci numbers is inefficient because the algorithm takes exponential time, whereas there exist dynamic programming solutions that take linear time.

To write a parallel version, we remark that the two recursive calls are completely independent: they do not depend on each other (neither uses a piece of data generated or written by another). We can therefore perform the recursive calls in parallel. In general, any two independent functions can be run in parallel. To indicate that two functions can be run in parallel, we use fork2().

Suppose that we wish to map an array to another by incrementing each element by one. We can write the code for a function map_incr that performs this computation serially.

This is not a good parallel algorithm but it is not difficult to give a parallel algorithm for incrementing an array. The code for such an algorithm is given below.

It is easy to see that this algorithm has O(n) work and $O(\log{n})$ span.

In the Fibonacci example, we started with a sequential algorithm and derived a parallel algorithm by annotating independent functions. It is also possible to go the other way and derive a sequential algorithm from a parallel one. As you have probably guessed this direction is easier, because all we have to do is remove the calls to the fork2 function. The sequential elision of our parallel Fibonacci code can be written by replacing the call to fork2() with a statement that performs the two calls (arguments of fork2()) sequentially as follows.

The sequential elision is often useful for debugging and for optimization. It is useful for debugging because it is usually easier to find bugs in sequential runs of parallel code than in parallel runs of the same code. It is useful in optimization because the sequentialized code helps us to isolate the purely algorithmic overheads that are introduced by parallelism. By isolating these costs, we can more effectively pinpoint inefficiencies in our code.

We defined fork-join programs as a subclass case of multithreaded programs. Let’s see more precisely how we can "map" a fork-join program to a multithreaded program. An our running example, let’s use the map_incr_rec, whose code is reproduced below.

Since, a fork-join program does not explicitly manipulate threads, it is not immediately clear what a thread refers to. To define threads, we can partition a fork-join computation into pieces of serial computations, each of which constitutes a thread. What we mean by a serial computation is a computation that runs serially and also that does not involve any synchronization with other threads except at the start and at the end. More specifically, for fork-join programs, we can define a piece of serial computation a thread, if it executes without performing parallel operations (fork2) except perhaps as its last action. When partitioning the computation into threads, it is important for threads to be maximal; technically a thread can be as small as a single instruction.

The Figure above illustrates the dag for an execution of map_incr_rec. We partition each invocation of this function into two threads labeled by "M" and "C" respectively. The threads labeled by $M\lbrack i,j \rbrack $ corresponds to the part of the invocation of map_incr_rec with arguments lo and hi set to $i$ and $j$ respectively; this first part corresponds to the part of execution up and including the fork2 or all of the function if this is a base case. The second corresponds to the "continuation" of the fork2, which is in this case includes no computation.

Based on this dag, we can create another dag, where each thread is replaced by the sequence of instructions that it represents. This would give us a picture similar to the dag we drew before for general multithreaded programs. Such a dag representation, where we represent each instruction by a vertex, gives us a direct way to calculate the work and span of the computation. If we want to calculate work and span ond the dag of threads, we can label each vertex with a weight that corresponds to the number of instruction in that thread.

Let’s observe a few properties of fork-join computations and their dags.

Suppose now we are given a computation dag and we wish to execute the dag by mapping each thread to one of the $P$ processor that is available on the hardware. To do this, we can use an online scheduling algorithm.

Recall that the $n^{th}$ Fibonnacci number is defined by the recurrence relation

with base cases

To write a parallel version, we remark that the two recursive calls are completely independent: they do not depend on each other (neither uses a piece of data generated or written by another). We can therefore perform the recursive calls in parallel. In general, any two independent functions can be run in parallel. To indicate that two functions can be run in parallel, we use async(). It is a requirement that each async takes place in the context of a finish. All async'ed computations that takes place in the context of a finish synchronize at that finish, i.e., finish completes only when all the async'ed computations complete. The important point about a finish is that it can serve as the synchronization point for an arbitrary number of async'ed computations. This is not quite useful in Fibonacci because there are only two computations to synchronize at a time, but it will be useful in our next example.

Recall our example for mapping an array to another by incrementing each element by one. We can write the code for a function map_incr that performs this computation serially.

Using async-finish, we can write to code parallel array increment by using only a single ‘finish` block and async’ ing all the parallel computations inside that finish.

It is helpful to compare this to fork-join, where we had to synchronize parallel branches in pairs. Here, we are able to synchronize them all at a single point.

The Figure above illustrates the dag for an execution of map_incr_rec. Each invocation of this function corresponds to a thread labeled by "M". The threads labeled by $M\lbrack i,j \rbrack $ corresponds to the part of the invocation of map_incr_rec_aux with arguments lo and hi set to $i$ and $j$ respectively. Note that all threads synchronize at the single finish node.

Based on this dag, we can create another dag, where each thread is replaced by the sequence of instructions that it represents. This would give us a picture similar to the dag we drew before for general multithreaded programs. Such a dag representation, where we represent each instruction by a vertex, gives us a direct way to calculate the work and span of the computation. If we want to calculate work and span on the dag of threads, we can label each vertex with a weight that corresponds to the number of instruction in that thread.

Using async-finish does not alter the asymptotic work and span of the computation compared to fork-join, which remain as O(n) work and $O(\log{n})$ span. In practice, however, the async-finish version creates fewer threads (by about a factor of two), which can make a difference.

Recall that the $n^{th}$ Fibonnacci number is defined by the recurrence relation

with base cases

We can write a parallel version of Fibonacci using futures as follows.

Note that this code is very similar to the fork-join version. In fact, we can directly map any use of fork-join parallelism to futures.

Much the same way that we represent fork-join and async-finish computations with a dag of threads, we can also represent future-based computations with a dag. The idea is to spawn a subdag for each future and for each force add an edge from the terminus of that subdag to the vertex that corresponds to the force. For example, for the Fibonacci function, the dags with futures are identical to those with fork-join. As we shall see, however, we can create more interesting dags with futures.

Recall our example for mapping an array to another by incrementing each element by one. We can write the code for a function map_incr that performs this computation serially.

We can write a parallel version of this code using futures by following the same approach that we did for Fibonacci. But let’s consider something slightly different. Suppose that the array is given to us an array of future values. We can write a serial version of map_incr for such an array as follows.

The function future_map_incr takes an array of futures of long, written long^ and returns an array of the same type. To compute each value of the array, the function force's the corresponding element of the source inside a future.

Futures enable expressing parallelism at a very fine level of granularity, at the level of individual data dependencies. This in turn allows parallelizing computations at a similarly finer grain, enabling a technique called pipelining.

The idea behind pipelining is to decompose a task $T$ into a sequence of smaller subtasks $T_0, \ldots, T_k$ to be performed in that order and overlap the computation of multiple tasks by starting another instance of the same kind of task as soon as the first subtask completes. As a result, if we have a collection of the same kinds of tasks that can be decomposed in the same way, we can have multiple of them "in flight" without having to wait for one to complete.

When computing sequentially, pipelining does not help performance because we can perform one computation at each time step. But when using parallel hardware, we can keep multiple processors busy by pipelining.

Suppose as an example, we have a sequence of tasks $T^0, T^1, T^2, T^3, T^4, T^5$ that we wish to compute. Suppose furthermore that these task depend on each other, that is a later task use the results of an earlier task. Thus it might seem that there is no parallelism that we can exploit. Upon further inspection, however, we may realize that we can partition each task $T_i$ into subtasks $T_0^i, T_1^i, T_2^i, T_3^i, T_4^i, T_5^i$ such that the subtask $j$ of task $i$ is used by the subtask $j+1$ of task $i+1$. We can then execute these tasks in parallel as shown below. Thus in the "steady state" where we have a large supply of these tasks, we can increase performance by a factor $5$, the number of subtasks that a task decomposes.

This idea of pipelining turns out to be quite important in some algorithms, leading sometimes to asymptotic improvements in run-time. It can, however, be painful to design the algorithm to take advantage of pipelining, especially if all we have available at our disposal are fork-join and async-finish parallelism, which require parallel computations to be independent. When using these constructs, we might therefore have to redesign the algorithm so that independent computations can be structurally separated and spawned in parallel. On the other hand, futures make it trivial to express pipelined algorithms because we can express data dependencies and ignore how exactly the individual computations may need to be executed in parallel. For example, in our hypothetical example, all we have to do is create a future for each sub-task, and force the relevant subtask as needed, leaving it to the scheduler to take advantage of the parallelism made available.

As a more concrete example, let’s go back to our array increment function and generalize to array map, and assume that we have another function mk_array that populates the contents of the array.

In this example, the $i^{th}$ element of the destination array depends only on the $i^{th}$ element of the source, thus as soon as that element is available, the function g might be invoked. This allows us to pipeline the execution of the functions mk_array and future_map_g.

While we shall not discuss this in detail, it is possible to improve the asymptotic complexity of certain algorithms by using futures and pipelining.

An important research question regarding futures is their scheduling. One challenge is primarily contention. When using futures, it is easy to create dag vertices, whose out-degree is non-constant. The question is how can such dag nodes can be represented to make sure that they don’t create a bottleneck, while also allowing the small-degree instance, which is the common case, to proceed efficiently. Another challenge is data locality. Async-finish and fork-join programs can be scheduled to exhibit good data locality, for example, using work stealing. With futures, this is more tricky. It can be shown for example, that even a single scheduling action can cause a large negative impact on the data locality of the computation.

In parallel programming, mutual exclusion problems do not have to arise. For example, if we program in a purely functional language extended with structured multithreading primitives such as fork-join and futures, programs remain purely functional and mutual-exclusion problems, and hence race conditions, do not arise. If we program in an imperative language, however, where memory is always a shared resource, even when it is not intended to be so, threads can easily share memory objects, even unintentionally, leading to race conditions. .Writing to the same location in parallel.

As in the example shows, separate threads are updating the value result but it might look like this is not a race condition because the update consists of an addition operation, which reads the value and then writes to i. The race condition might be easier to see if we expand out the applications of the += operator.

When written in this way, it is clear that these two parallel threads are not independent: they both read result and write to result. Thus the outcome depends on the order in which these reads and writes are performed, as shown in the next example.

Since mutual exclusion is a common problem in computer science, many hardware systems provide specific synchronization operations that can help solve instances of the problem. These operations may allow, for example, testing the contents of a (machine) word then modifying it, perhaps by swapping it with another word. Such operations are sometimes called atomic read-modify-write or RMW, for short, operations.

A handful of different RMW operations have been proposed. They include operations such as load-link/store-conditional, fetch-and-add, and compare-and-swap. They typically take the memory location x, and a value v and replace the value of stored at x with f(x,v). For example, the fetch-and-add operation takes the location x and the increment-amount, and atomically increments the value at that location by the specified amount, i.e., f(x,v) = *x + v.

The compare-and-swap operation takes the location x and takes a pair of values (a,b) as the second argument, and stores b into x if the value in x is a, i.e., f(x,(a,b)) = if *x = a then b else a; the operation returns a Boolean indicating whether the operation successfully stored a new value in x. The operation "compare-and-swap" is a reasonably powerful synchronization operation: it can be used by arbitrarily many threads to agree (reach consensus) on a value. This instruction therefore is frequently provided by modern parallel architectures such as Intel’s X86.

In C$++$, the atomic class can be used to perform synchronization. Objects of this type are guarantee to be free of race conditions; and in fact, in C++, they are the only objects that are guaranteed to be free from race conditions. The contents of an atomic object can be accessed by load operations, updated by store operation, and also updated by compare_exchange_weak and compare_exchange_strong operations, the latter of which implement the compare-and-swap operation.

The key operation that help with race conditions is the compare-and-exchange operation.

As another example use of the atomic class, recall our Fibonacci example with the race condition. In that example, race condition arises because of concurrent writes to the result variable. We can eliminate this kind of race condition by using different memory locations, or by using an atomic class and using a compare_exchange_strong operation.

The example above illustrates a typical use of the compare-and-swap operation. In this particular example, we can probably prove our code is correct. But this is not always as easy due to a problem called the "ABA problem."

While reasonably powerful, compare-and-swap suffers from the so-called ABA problem. To see this consider the following scenario where a shared variable result is update by multiple threads in parallel: a thread, say $T$, reads the result and stores its current value, say 2, in current. In the mean time some other thread also reads result and performs some operations on it, setting it back to 2 after it is done. Now, thread $T$ takes its turn again and attempts to store a new value into result by using 2 as the old value and being successful in doing so, because the value stored in result appears to have not changed. The trouble is that the value has actually changed and has been changed back to the value that it used to be. Thus, compare-and-swap was not able to detect this change because it only relies on a simple shallow notion of equality. If for example, the value stored in result was a pointer, the fact that the pointer remains the same does not mean that values accessible from the pointer has not been modified; if for example, the pointer led to a tree structure, an update deep in the tree could leave the pointer unchanged, even though the tree has changed.

This problem is called the ABA problem, because it involves cycling the atomic memory between the three values $A$, $B$, and again $A$). The ABA problem is an important limitation of compare-and-swap: the operation itself is not atomic but is able to behave as if it is atomic if it can be ensured that the equality test of the subject memory cell suffices for correctness.

In the example below, ABA problem may happen (if the counter is incremented and decremented again in between a load and a store) but it is impossible to observe because it is harmless. If however, the compare-and-swap was on a memory object with references, the ABA problem could have had observable effects.

The ABA problem can be exploited to give seemingly correct implementations that are in fact incorrect. To reduce the changes of bugs due to the ABA problem, memory objects subject to compare-and-swap are usually tagged with an additional field that counts the number of updates. This solves the basic problem but only up to a point because the counter itself can also wrap around. The load-link/store-conditional operation solves this problem by performing the write only if the memory location has not been updated since the last read (load) but its practical implementations are hard to come by.

Let’s start by downloading the PASL sources. The PASL sources that we are going to use are part of a branch that we created specifically for this course. You can access the sources either via the tarball linked by the github webpage or, if you have git, via the command below.

You can skip this section if you are using a computer already setup by us or you have installed an image file containing our software. To skip this part and use installed binaries, see the heading "Starting with installed binaries", below.

At this point, you have either installed all the necessary software to work with PASL or these are installed for you. In either case, make sure that your PATH variable makes the software visible. For setting up your PATH variable on andrew.cmu domain, see below.

PASL comes equipped with a Makefile that can generate several different kinds of executables. These different kinds of executables and how they can be generated is described below for a benchmark program pgm.

To speed up the build process, add to the make command the option -j (e.g., make -j pgm.opt). This option enables make to parallelize the build process. Note that, if the build fails, the error messages that are printed to the terminal may be somewhat garbled. As such, it is better to use -j only if after the debugging process is complete.

We are going to start our experimentation with three different instances of the same program, namely bench. This program serves as a "driver" for the benchmarks that we have implemented. These implementations are good parallel codes that we expect to deliver good performance. We first build the baseline version.

The file extension .baseline means that every benchmark in the binary uses the sequential-baseline version of the specified algorithm.

We can now run the baseline for one of our benchmarks, say Fibonacci by using the -bench argument to specify the benchmark and the -n argument to specify the input value for the Fibonacci function.

On our machine, the output of this run is the following.

The three lines above provide useful information about the run.

The .elision extension means that parallel algorithms (not sequential baseline algorithms) are compiled. However, all instances of fork2() are erased as described in an earlier chapter.

The run time of the sequential elision in this case is similar to the run time of the sequential baseline because the two are similar codes. However, for most other algorithms, the baseline will typically be at least a little faster.

The .opt extension means that the program is compiled with full support for parallel execution. Unless specified otherwise, however, the parallel binary uses just one processor.

The output of this program is similar to the output of the previous two programs.

Because our machine has 40 processors, we can run the same application using all available processors. Before running this command, please adjust the -proc option to match the number of cores that your machine has. Note that you can use any number of cores up to the number you have available. You can use nproc or lscpu to determine the number of cores your machine has.

We see from the output of the 40-processor run that our program ran faster than the sequential runs. Moreover, the utilization field tells us that approximately 86% of the total time spent by the 40 processors was spent performing useful work, not idling.

We may ask at this point: What is the improvement that we just observed from the parallel run of our program? One common way to answer this question is to measure the "speedup".

The speedup at a given number of processors is a good starting point on the way to evaluating the scalability of the implementation of a parallel algorithm. The next step typically involves considering speedups taken from varying numbers of processors available to the program. The data collected from such a speedup experiment yields a speedup curve, which is a curve that plots the trend of the speedup as the number of processors increases. The shape of the speedup curve provides valuable clues for performance and possibly for tuning: a flattening curve suggests lack of parallelism; a curve that arcs up and then downward suggests that processors may be wasting time by accessing a shared resource in an inefficient manner (e.g., false sharing); a speedup curve with a constant slope indicates at least some scaling.

The 29x speedup that we just calculated for our Fibonacci benchmark was a little dissapointing, and the 86% processor utilization of the run left 14% utilization for improvement. We should be suspicious that, although seemingly large, the problem size that we chose, that is, $n = 39$, was probably a little too small to yield enough work to keep all the processors well fed. To put this hunch to the test, let us examine the utilization of the processors in our system. We need to first build a binary that collects and outputs logging data.

We run the program with the new binary in the same fashion as before.

The output looks something like the following.

We need to explain what the new fields mean.

Each of these fields can be useful for tracking down inefficiencies. The number of freshly allocated threads can be a strong indicator because in C++ thread allocation costs can sometimes add up to a significant cost. In the present case, however, none of the new values shown above are highly suspicious, considering that there are all at most in the thousands.

Since we have not yet found the problem, let us look at the visualization of the processor utilization using our pview tool. To get the necessary logging data, we need to run our program again, this time passing the argument --pview.

When the run completes, a binary log file named LOG_BIN should be generated in the current directory. Every time we run with --pview this binary file is overwritten. To see the visualization of the log data, we call the visualizer tool from the same directory.

The output we see on our 40-processor machine is shown in the Figure below. The window shows one bar per processor. Time goes from left to right. Idle time is represented by red and time spent busy with work by grey. You can zoom in any part of the plot by clicking on the region with the mouse. To reset to the original plot, press the space bar. From the visualization, we can see that most of the time, particularly in the middle, all of the processors keep busy. However, there is a lot of idle time in the beginning and end of the run. This pattern suggests that there just is not enough parallelism in the early and late stages of our Fibonacci computation.

We are pretty sure that or Fibonacci program is not scaling as well is it could. But poor scaling on one particular input for $n$ does not necessarily mean there is a problem with the scalability our parallel Fibonacci program in general. What is important is to know more precisely what it is that we want our Fibonacci program to achieve. To this end, let us consider a distinction that is important in high-performance computing: the distinction between strong and weak scaling. So far, we have been studying the strong-scaling profile of the computation of the $39^{th}$ Fibonacci number. In general, strong scaling concerns how the run time varies with the number of processors for a fixed problem size. Sometimes strong scaling is either too ambitious, owing to hardware limitations, or not necessary, because the programmer is happy to live with a looser notion of scaling, namely weak scaling. In weak scaling, the programmer considers a fixed-size problem per processor. We are going to consider something similar to weak scaling. In the Figure below, we have a plot showing how processor utilization varies with the input size. The situation dramatically improves from 12% idle time for the $39^{th}$ Fibonacci number down to 5% idle time for the $41^{st}$ and finally to 1% for the $45^{th}$. At just 1% idle time, the utilization is excellent.

The scenario that we just observed is typical of multicore systems. For computations that perform relatively little work, such as the computation of the $39^{th}$ Fibonacci number, properties that are specific to the hardware, OS, and PASL load-balancing algorithm can noticeably limit processor utilization. For computations that perform lots of highly parallel work, such limitations are barely noticeable, because processors spend most of their time performing useful work. Let us return to the largest Fibonacci instance that we considered, namely the computation of the $45^{th}$ Fibonacci number, and consider its utilization plot.

The utilization plot is shown in the Figure below. Compared the to utilization plot we saw in the Figure above for n=39, the red regions are much less prominent overall and the idle regions at the beginning and end are barely noticeable.

We have seen in this lab how to build, run, and evaluate our parallel programs. Concepts that we have seen, such as speedup curves, are going to be useful for evaluating the scalability of our future solutions. Strong scaling is the gold standard for a parallel implementation. But as we have seen, weak scaling is a more realistic target in most cases.

It is common for a parallel algorithm to be asymptotically and/or observably work inefficient but it is often possible to improve work efficiency by observing that work efficiency increases with parallelism and can thus be controlled by limiting it.

For example, a parallel algorithm that performs linear work and has logarithmic span leads to average parallelism in the orders of thousands with the small input size of one million. For such a small problem size, we usually would not need to employ thousands of processors. It would be sufficient to limit the parallelism so as to feed tens of processors and as a result reduce impact of excess parallelism on work efficiency.

In many parallel algorithms such as the algorithms based on divide-and-conquer, there is a simple way to achieve this goal: switch from parallel to sequential algorithm when the problem size falls below a certain threshold. This technique is sometimes called coarsening or granularity control.

But which code should we switch to: one idea is to simply switch to the sequential elision, which we always have available in PASL. If, however, the parallel algorithm is asymptotically work inefficient, this would be ineffective. In such cases, we can specify a separate sequential algorithm for small instances.

Optimizing the practical efficiency of a parallel algorithm by controlling its parallelism is sometimes called optimization, sometimes it is called performance engineering, and sometimes performance tuning or simply tuning. In the rest of this document, we use the term "tuning."

The basic idea behind coarsening or granularity control is to revert to a fast serial algorithm when the input size falls below a certain threshold. To determine the optimal threshold, we can simply perform a search by running the code with different threshold settings.

While this approach can help find the right threshold on the particular machine that we performed the search, there is no guarantee that the same threshold would work on another machine. In fact, there are examples in the literature that show that such optimizations are not portable, i.e., a piece of code optimized for a particular architecture may behave poorly on another.

In the general case, determining the right threshold is even more difficult. To see the difficulty consider a generic (polymorphic), higher-order function such as map that takes a sequence and a function and applies the function to the sequence. The problem is that the threshold depends both on the type of the elements of the sequence and the function to be mapped over the sequence. For example, if each element itself is a sequence (the sequence is nested), the threshold can be relatively small. If, however, the elements are integers, then the threshold will likely need to be relatively large. This makes it difficult to determine the threshold because it depends on arguments that are unknown at compile time. Essentially the same argument applies to the function being mapped over the sequence: if the function is expensive, then the threshold can be relatively small, but otherwise it will need to be relatively large.

As we describe in this chapter, it is sometimes possible to determine the threshold completely automatically.

Our automatic granularity-control technique requires assistance from the application programmer: for each parallel region of the program, the programmer must annotate the region with a complexity function, which is simply a C++ function that returns the asymptotic work cost associated with running the associated region of code.

More generally, suppose that we know that a given algorithm has work cost of $W = n + \log n$. Although it would be fine to assign to this algorithm exactly $W$, we could just as well assign to the algorithm the cost $n$, because the second term is dominated by the first. In other words, when expressing work costs, we only need to be precise up to the asymptotic complexity class of the work.

In PASL, a controlled statement, or cstmt, is an annotation in the program text that activates automatic granularity control for a specified region of code. In particular, a controlled statement behaves as a C++ statement that has the special ability to choose on the fly whether or not the computation rooted at the body of the statement spawns parallel threads. To support such automatic granularity control PASL uses a prediction algorithm to map the asymptotic work cost (as returned by the complexity function) to actual processor cycles. When the predicted processor cycles of a particular instance of the controlled statement falls below a threshold (determined automatically for the specific machine), then that instance is sequentialized, by turning off the ability to spawn parallel threads for the execution of that instance. If the predicted processor cycle count is higher than the threshold, then the statement instance is executed in parallel.

In other words, the reader can think of a controlled statement as a statement that executes in parallel when the benefits of parallel execution far outweigh its cost and that executes sequentially in a way similar to the sequential elision of the body of the controlled statement would if the cost of parallelism exceeds its benefits. We note that while the sequential exection is similar to a sequential elision, it is not exactly the same, because every call to fork2 must check whether it should create parallel threads or run sequentially. Thus the execution may differ from the sequential elision in terms of performance but not in terms of behavior or semantics.

The controlled statement takes three arguments, whose requirements are specified below, and returns nothing (i.e., void). The effectiveness of the granularity controller may be compromised if any of the requirements are not met.

When the controlled statement chooses sequential evaluation for its body the effect is similar to the effect where in the code above the input size falls below the threshold size: the body and the recursion tree rooted there is sequentialized. When the controlled statement chooses parallel evaluation, the calls to fork2() create parallel threads.

Let us add one more component to our granularity-control toolkit: the parallel-for from. By using this loop construct, we can avoid having to explicitly express recursion-trees over and over again. For example, the following function performs the same computation as the example function we defined in the first lecture. Only, this function is much more compact and readable. Moreover, this code takes advantage of our automatic granularity control, also by replacing the parallel-for with a serial-for.

Underneath, the parallel-for loop uses a divide-and-conquer routine whose structure is similar to the structure of the divide-and-conquer routine of our map_incr_rec. Because the parallel-for loop generates the log-height recursion tree, the map_incr routine just above has the same span as the map_incr routine that we defined earlier: $\log n$, where $n$ is the size of the input array.

Notice that the code above specifies no complexity function. The reason is that this particular instance of the parallel-for loop implicitly defines a complexity function. The implicit complexity function reports a linear-time cost for any given range of the iteration space of the loop. In other words, the implicit complexity function assumes that per iteration the body of the loop performs a constant amount of work. Of course, this assumption does not hold in general. If we want to specify explicitly the complexity function, we can use the form shown in the example below. The complexity function is passed to the parallel-for loop as the fourth argument. The complexity function takes as argument the range [lo, hi). In this case, the complexity is linear in the number of iterations. The function simply returns the number of iterations as the complexity.

The following code snippet shows a more interesting case for the complexity function. In this case, we are performing a multiplication of a dense matrix by a dense vector. The outer loop iterates over the rows of the matrix. The complexity function in this case gives to each of these row-wise iterations a cost in proportion to the number of scalars in each column.

While parallel matrix multiplication delivers excellent speedups, this is not common for many other algorithms on modern multicore machines where many computations can quickly become limited by the availability of bandwidth. Matrix multiplication does not suffer as much from the memory-bandwidth limitations because it performs significant work per memory operation: it touches $\Theta(n^2)$ memory cells, while performing nearly $\Theta(n^3)$ work.

The key components of our array data structure, sparray, are shown by the code snippet below. An sparray can store 64-bit words only; in particular, they are monomorphic and fixed to values of type long. Generalizing sparray to store values of arbitrary types can be achieved by use of C++ templates. We stick to monomorphic arrays here to simplify the presentation.

The class sparray provides two constructors. The first one takes in the size of the array (set to 0 by default) and allocates an unitialized array of the specified size (nullptr if size is 0). The second constructor takes in a list specified by curly braces and allocates an array with the same size. Since the argument to this constructor must be specified explicitly in the program, its size is constant by definition.

The cost model guaranteed by our implementation of parallel array is as follows:

Arrays can be allocated by specifying the size of the array.

Just after creation, the array contents consist of uninitialized bits. We use this convention because the programmer needs flexibility to decide the parallelization strategy to initialize the contents. Internally, the sparray class consists of a size field and a pointer to the first item in the array. The contents of the array are heap allocated (automatically) by constructor of the sparray class. Deallocation occurs when the array’s destructor is called. The destructor can be called by the programmer or by run-time system (of C++) if an object storing the array is destructed. Since C++ destructs (stack allocated) variables that go out of scope when a function returns, we can combine the stack discipline with heap-allocated arrays to manage the deallocation of arrays mostly automatically. We give several examples of this automatic deallocation scheme below.

It is safe to take a pointer to a cell in the array, when the array itself is still in scope. For example, in the code below, the contents of the array are used strictly when the array is in scope.

We are going to see that we can rely on cleaner conventions for passing to functions references on arrays.

If you are familiar with C++ container libraries, such as STL, this aspect of our array implementation, namely the calling conventions, may be unfamiliar: our arrays cannot be passed by value. We forbid passing by value because passing by value implies creating a fresh copy for each array being passed to or returned by a function. Of course, sometimes we really need to copy an array. In this case, we choose to copy the array explicitly, so that it is obvious where in our code we are paying a linear-time cost for copying out the contents. We will return to the issue of copying later.

Returning an array is straightforward: we take advantage of a feature of modern C++11 which automatically detects when it is safe to move a structure by a constant-time pointer swap. Code of the following form is perfectly legal, even though we disabled the copy constructor of sparray, because the compiler is able to transfer ownership of the array to the caller of the function. Moreover, the transfer is guaranteed to be constant work — not linear like a copy would take. The return is fast, because internally all that happens is that a couple words are being exchanged. Such "move on return" is achieved by the "move-assignment operator" of sparray class.

Although it is perfectly fine to assign to an array variable the contents of a given array, what happens may be surprising to those who know the usual conventions of C++11 container libraries. Consider the following program.

The reason we use this semantics for assignment is that the assignment takes constant time. Later, we are going to see that we can efficiently copy items out of an array. But for reasons we already discussed, the copy operation is going to be explicit.

Exercise: duplicating items in parallel

A tabulation is a parallel operation which creates a new array of a given size and initializes the contents according to a given "generator function". The call tabulate(g, n) allocates an array of length n and assigns to each valid index in the array i the value returned by g(i).

Tabulations can be used to generate sequences according to a specified formula.

Copying an array can be expressed as a tabulation.

The implementation of tabulate is a straightforward application of the parallel-for loop.

Note that the work and span of the generator function depends on the generator function passed as an argument to the tabulation. Let us first analyze for the simple case, where the generator function takes constant work (and hence, constant span). In this case, it should be clear that a tabulation should take work linear in the size of the array. The reason is that the only work performed by the body of the loop is performed by the constant-time generator function. Since the loop itself performs as many iterations as positions in the array, the work cost is indeed linear in the size of the array. The span cost of the tabulation is the sum of two quantitles: the span taken by the loop and the maximum value of the spans taken by the applications of the generator function. Recall that we saw before that the span cost of a parallel-for loop with $n$ iterations is $\log n$. The maximum of the spans of the generator applications is a constant. Therefore, we can conclude that, in this case, the span cost is logarithmic in the size of the array.

The story is only a little more complicated when we generalize to consider non-constant time generator functions. Let $W(\mathtt{g}(i))$ denote the work performed by an application of the generator function to a specified value $i$. Similarly, let $S(\mathtt{g}(i))$ denote the span. Then the tabulation takes work

and span

We just observed that each application of our tabulate operation can have different work and span cost depending on the selection of the generator function. Pause for a moment and consider how this observation could impact our granularity-control scheme. Consider, in particular, the way that the tabulate function uses its granularity-controller object, tabulate_contr. This one controller object is shared by every call site of tabulate().

The problem is that all of the profiling data that the granularity controller collects at run time is lumped together, even though each generator function that is passed to the tabulate function can have completely different performance characteristics. The threshold that is best for one generator function is not necessarily good for another generator function. For this reason, there must be one distinct granularity-control object for each generator function that is passed to tabulate. For this reason, we refine our solution from the one above to the one below, which relies on C++ template programming to effectively key accesses to the granularity controller object by the type of the generator function.

To be more precise, the use of the template parameter in the class tabulate_controller ensures that each generator function in the program that is passed to tabulate() gets its own unique instance of the controller object. The rules of the template system regarding static class members that appear in templated classes ensure this behavior. Although it is not essential for our purposes to have a precise understanding of the template system, it is useful to know that the template system provides us with the exact mechanism that we need to properly separate granularity controllers of distinct instances of higher-order functions, such as tabulation.

A reduction is an operation which combines a given set of values according to a specified identity element and a specified associative combining operator. Let $S$ denote a set. Recall from algebra that an associative combining operator is any binary operator $\oplus$ such that, for any three items $x,y,z \in S$, the following holds.

An element $\mathbf{I} \in S$ is an identity element if for any $x \in S$ the following holds.

This algebraic structure consisting of $(S, \oplus, \mathbf{I})$ is called a monoid and is particularly worth knowing because this structure is a common pattern in parallel computing.

The identity element is important because we are working with sequences: having a base element is essential for dealing with empty sequences. For example, what should the sum of the empty sequence? More interestingly, what should be the maximum (or minimum) element of an empty sequence? The identity element specifies this behavior.

What about the associativity of $\oplus$? Why does associativity matter? Suppose we are given the sequence $ [ a_0, a_1, \ldots, a_n ] $. The serial reduction of this sequence always computes the expression $(a_0 \oplus a_1 \oplus \ldots \oplus a_n)$. However, when the reduction is performed in parallel, the expression computed by the reduction could be $( (a_0 \oplus a_1 \oplus a_2 \oplus a_3) \oplus (a_4 \oplus a_5) \oplus \ldots \oplus (a_{n-1} \oplus a_n) )$ or $( (a_0 \oplus a_1 \oplus a_2) \oplus (a_3 \oplus a_4 \oplus a_5) \oplus \ldots \oplus (a_{n-1} \oplus a_n) )$. In general, the exact placement of the parentheses in the parallel computation depends on the way that the parallel algorithm decomposes the problem. Associativity gives the parallel algorithm the flexibility to choose an efficient order of evaluation and still get the same result in the end. The flexibility to choose the decomposition of the problem is exploited by efficient parallel algorithms, for reasons that should be clear by now. In summary, associativity is a key building block to the solution of many problems in parallel algorithms.

Now that we have monoids for describing a generic method for combining two items, we can consider a generic method for combining many items in parallel. Once we have this ability, we will see that we can solve the remaining problems from last homework by simply plugging the appropriate monoids into our generic operator, reduce. The interface of this operator in our framework is specified below. The first parameter corresponds to $\oplus$, the second to the identity element, and the third to the sequence to be processed.

We can solve our first problem by plugging integer plus as $\oplus$ and 0 as $\mathbf{I}$.

We can solve our second problem in a similar fashion. Note that in this case, since we know that the input sequence is nonempty, we can pass the first item of the sequence as the identity element. What could we do if we instead wanted a solution that can deal with zero-length sequences? What identity element might make sense in that case? Why?

Like the tabulate function, reduce is a higher-order function. Just like any other higher-order function, the work and span costs have to account for the cost of the client-supplied function, which is in this case, the associative combining operator.

A scan is an iterated reduction that is typically expressed in one of two forms: inclusive and exclusive. The inclusive form maps a given sequence $ [ x_0, x_1, x_2, \ldots, x_{n-1} ] $ to $ [ x_0, x_0 \oplus x_1, x_0 \oplus x_1 \oplus x_2, \ldots, x_0 \oplus x_1 \oplus \ldots \oplus x_{n-1} ] $.

The exclusive form maps a given sequence $ [ x_0, x_1, x_2, \ldots, x_{n-1} ] $ to $ [ \mathbf{I}, x_0, x_0 \oplus x_1, x_0 \oplus x_1 \oplus x_2, \ldots, x_0 \oplus x_1 \oplus \ldots \oplus x_{n-2} ] $. For convenience, we extend the result of the exclusive form with the total $ x_0 \oplus \ldots \oplus x_{n-1} $.

Scan has applications in many parallel algorithms. To name just a few, scan has been used to implement radix sort, search for regular expressions, dynamically allocate processors, evaluate polynomials, etc. Suffice to say, scan is important and worth knowing about because scan is a key component of so many efficient parallel algorithms. In this course, we are going to study a few more applications not in this list.

The expansions shown above suggest the following sequential algorithm.

If we just blindly follow the specification above, we might be tempted to try the solution below.

Consider that our sequential algorithm takes linear time in the size of the input array. As such, finding a work-efficient parallel solution means finding a solution that also takes linear work in the size of the input array. The problem is that our parallel algorithm takes quadratic work: it is not even asymptotically work efficient! Even worse, the algorithm performs a lot of redundant work.

Can we do better? Yes, in fact, there exist solutions that take, in the size of the input, both linear time and logarithmic span, assuming that the given associative operator takes constant time. It might be worth pausing for a moment to consider this fact, because the specification of scan may at first look like it would resist a solution that is both highly parallel and work efficient.

The remaining operations that we are going to consider are useful for writing more succinct code and for expressing special cases where certain optimizations are possible. All of the the operations that are presented in this section are derived forms of tabulate, reduce, and scan.

The quicksort algorithm for sorting an array (sequence) of elements is known to be a very efficient sequential sorting algorithm. A natural question thus is whether quicksort is similarly effective as a parallel algorithm?

Let us first convince ourselves, at least informally, that quicksort is actually a good parallel algorithm. But first, what do we mean by "parallel quicksort." Chances are that you think of quicksort as an algorithm that, given an array, starts by reorganizing the elements in the array around a randomly selected pivot by using an in-place partitioning algorithm, and then sorts the two parts of the array to the left and the right of the array recursively.

While this implementation of the quicksort algorithm is not immediately parallel, it can be parallelized. Note that the recursive calls are naturally independent. So we really ought to focus on the partitioning algorithm. There is a rather simple way to do such a partition in parallel by performing three filter calls on the input array, one for picking the elements less that the pivot, one for picking the elements equal to the pivot, and another one for picking the elements greater than the pivot. This algorithm can be described as follows.

Now that we have a parallel algorithm, we can check whether it is a good algorithm or not. Recall that a good parallel algorithm is one that has the following three characteristics

Let us first convince ourselves that Quicksort is a highly parallel algorithm. Observe that