OpenCL book PDF

1.3 ParallelComputing (Software)

Up until this point, hardware architectures that involve performing one task using numerous processors had been the main focus. This section will focus on the method for parallel programming for the discussed hardware architectures.

Sequential vs Parallel Processing

Take a look at the pseudocode below.

List1.1: Pseudocode for parallel processing

  resultA	= task_a(i); 
  resultB	= task_b(i);
  resultC	= task_c(i);
  resultD	= task_d(i); 
  resultAll	+= resultA + resultB + resultC + resultD; 

Executing the above code on a single-core, single-CPU processor (SISD system), the 4 tasks would be run sequentially in the order task_a, task_b, task_c, task_d, and the returned values are then summed up. This is then repeated N times, incrementing i after each iteration. This type of method is called Sequential processing. Running the code with N=4 is shown on the left hand side of Figure 1.3.

This type of code can benefit from parallelization. If this code is compiled without adding any options for optimization on a Dual Core system, the program would run sequentially on one core. In this scenario, the other core has nothing to do, so it becomes idle. This is clearly inefficient, so the intuitive thing to do here is to split the task into 2 subtasks, and run each subtasks on each core. This is the basis of parallel programming. As shown on Figure 1.3, the task is split up into 2 such that each subtask runs the loop N/2 times, and then performs addition at the end.

Figure 1.3: Parallel processing example

For actual implementation of parallel programs, it is necessary to follow the following steps.

1. Analyze the dependency within the data structures or within the processes, etc, in order to decide which sections can be executed in parallel.

2. Decide on the best algorithm to execute the code over multiple processors

3. Rewrite code using frameworks such as Message Passing Interface (MPI), OpenMP, or OpenCL.

In the past, these skills were required only by a handful of engineers, but since multi-core processors are becoming more common, the use of these distributed processing techniques are becoming more necessary. The following sections will introduce basic concepts required to implement parallel processes.

Where to Parallelize?

During the planning phase of parallel programs, certain existing laws must be taken into account. The first law states that if a program spends y% of the time running code that cannot be executed in parallel, the expected speed up from parallelizing is at best a 1/y fold improvement. The law can be proven as follows. Assume Ts to represent the time required to run the portion of the code that cannot be parallelized, and Tp to represent the time required to run the portion of the code that can benefit from parallelization. Running the program with 1 processor, the processing time is:

T(1) = Ts + Tp

The processing time when using N processors is:

T(N) = Ts + Tp/N

Using y to represent the proportion of the time spent running code that cannot be parallelized, the speedup S achieved is:

S = T(1)/T(N) = T(1)/(Ts + Tp/N) = 1/(y+(1-y)/N)

Taking the limit as N goes to infinity, the most speedup that can be achieved is S=1/y. This law is known as Amdahl's Law .

For a more concrete example, assume that a sequential code is being rewritten to run on a Quad Core CPU. Ideally, a 4-time speedup is achieved. However, as the law states, the speedup is limited by the portion of the code that must be run sequentially. Figure 1.4 shows the 2 cases where the proportion of the code that cannot be run in parallel (y) is 10%, and 50%. Even without taking overhead into account, the figure shows a difference of 3x and 1.3x speedup depending on y.

Figure 1.4: Amdahl's Law Example

This problem becomes striking as the number of processors is increased. For a common 2-way server that uses Intel's Xeon 5500 Series CPUs which supports hyper-threading, the OS sees 16 cores. GPUs such as NVIDIA's Tesla can have more 200 cores. Figure 1.5 shows the speedup achieved as a function of the sequentially processed percentage y and the number of cores. The graph clearly shows the importance of reducing the amount of sequentially processed portions, especially as the number of cores is increased. This also implies that the effort used for parallelizing a portion of the code that does not take up a good portion of the whole process may be in vain. In summary, it is more important to reduce serially processed portions rather than parallelizing a small chunk of the code.

Figure 1.5: Result of scaling for different percentage of tasks that cannot be parallelized

While Amdahl's law gives a rather pessimistic impression of parallel processing, the Gustafson Law provides a more optimistic view. This law states that as the program size increases, the fraction of the program that can be run in parallel also increases, thereby decreasing the portion that must be performed sequentially. Recall the previously stated equation:

T(N) = Ts + Tp/N

Gustafson states that of the changes in Ts, Tp is directly proportional to the program size, and that Tp grows faster than Ts. For example, assume a program where a portion that must be run sequentially is limited to initialization and closing processes, and all other processes can be performed in parallel. By increasing the amount of data to process by the program, it is apparent that Gustafson's law holds true. In other words, Gustafson's law shows that in order to efficiently execute code over multiple processors, a large-scale processing must take place. Development of parallel programming requires close attention to these 2 laws.

Types of Parallelism

After scrutinizing the algorithm and deciding where to parallelize, the next step is to decide on the type of parallelism to use.

Parallel processing requires the splitting up of the data to be handled, and/or the process itself. Refer back to the code on List 1.1. This time, assume the usage of a Quad-core CPU to run 4 processes at once. An intuitive approach is to let each processor perform N/4 iteration of the loop, but since there are 4 tasks within the loop, it also makes just as much sense to run each of these tasks in each processor. The former method is called "Data Parallel", and the latter method is called "Task Parallel" (Figure 1.6).

Figure 1.6: Data Parallel vs Task Parallel

Data Parallel

Main characteristics of data parallel method is that the programming is relatively simple since multiple processors are all running the same program, and that all processors finish their task at around the same time. This method is efficient when the dependency between the data being processed by each processor is minimal. For example, vector addition can benefit greatly from this method. As illustrated in Figure 1.7, the addition at each index can be performed completely independently of each other. For this operation, the number of processors is directly proportional to the speedup that may be achieved if overhead from parallelization can be ignored. Another, more concrete example where this method can be applied is in image processing. The pixels can be split up into blocks, and each of these blocks can be filtered in parallel by each processor.

Figure 1.7: Vector Addition

Task Parallel

The main characteristic of the task parallel method is that each processor executes different commands. This increases the programming difficulty when compared to the data parallel method. Since the processing time may vary depending on how the task is split up, it is actually not suited for the example shown in Figure 1.6. Also, since task_a and task_c are doing nothing until task_b and task_d finishes, the processor utilization is decreased. Task parallel method requires a way of balancing the tasks to take full advantage of all the cores. One way is to implement a load balancing function on one of the processors. In the example in Figure 1.8, the load balancer maintains a task queue, and assigns a task to a processor that has finished its previous task.

Figure 1.8: Load Balancing

Another method for task parallelism is known as pipelining. Pipelining is usually in reference to the "instruction pipeline" where multiple instructions, such as instruction decoding, arithmetic operation, and register fetch, are executed in an overlapped fashion over multiple stages. This concept can be used in parallel programming as well.

Figure 1.9 shows a case where each processor is given its own task type that it specializes in. In this example, the start of each task set is shifted in the time domain, such that task_b, task_c, task_d takes in the output of task_a, task_b, task_c as an input. The data moves as a stream across the processors. This method is not suited for the case where only one set of tasks is performed, but can be effective when processing, for example, videos, where processing frames are taken as inputs one after another.

Figure 1.9: Pipelining

Hardware Dependencies

When porting sequential code to parallel code, the hardware must be decided wisely. Programs usually have sections suited for data parallelism as well as for task parallelism, and the hardware is usually suited for one or the other.

For example, the GPU is suited for data parallel algorithms, due to the existence of many cores. However, since the GPU (at present) is not capable of performing different tasks in parallel, the Cell/B.E. is more suited for performing task parallel algorithms, since its 8 cores are capable of working independently of each other.

OpenCL allows the same code to be executed on either platforms, but since it cannot change the nature of the hardware, the hardware and the parallelization method must be chosen wisely.

Implementing a Parallel Program

After deciding on the parallelization method, the next step is the implementation. In decreasing order of user involvement:

1. Write parallel code using the operating system's functions.

2. Use a parallelization framework for program porting.

3. Use an automatic-parallelization compiler.

This section will explore the different methods.

Parallelism using the OS System Calls

Implementing parallel programs using the OS system call requires, at minimum, a call to execute and close a program, and some way of transferring data between the executed programs. If this is done on a cluster system, the data transfer between programs is performed using network transfer APIs such as the socket system call, but this is commonly done using a framework instead.

For performing parallel instructions performed within the processor itself, however, the OS system call may be used instead of the framework. The code can be further broken down into "parallel processes" and "parallel threads" to be run on the processor. The difference between processes and threads are as follows.

A process is an executing program given its own address space by the operating system. In general, the operating system performs execution, closing, and interruption within these process units, making sure each of these processes' distinct resources do not interfere with each other. Data transfer between programs may be performed by a system call to the operating system. For example, UNIX provides a system call shmget() that allocates shared memory that can be accessed by different processes.

A thread is a subset of a process that is executed multiple times within the program. These threads share the same address space as the processes. In general, since these threads execute in the same memory space, the overhead from starting and switching is much smaller than when compared to processes. In general, the operating system provides an API to create and manage these threads. For example, UNIX provides a library called Pthreads, which is a POSIX-approved thread protocol. POSIX is a standard for API specified by IEEE.

Whether to use parallel processes or parallel threads within a processor depends on the case, but in general, parallel threads are used if the goal is speed optimization, due to the light overhead.

List 1.2 shows an example where each member of an array is incremented using multithreading.

List 1.2: pthread example

#include <stdio.h>
#include <stdlib.h>
#include <pthread.h>
#define TH_NUM 4
#define N 100
static void *thread_increment(void *array)
int i;
int *iptr;
iptr = (int *)array;
for(i=0;i< N / TH_NUM;i++) iptr[i] += 1;
return NULL;
int main(void)
int i;
pthread_t thread[TH_NUM];
int array[N];
/* initialize array*/
array[i] = i;
/* Start parallel process */
if (pthread_create(&thread[i], NULL, thread_increment, array + i * N / TH_NUM) != 0)
return 1;
/* Synchronize threads*/
if (pthread_join(thread[i], NULL) != 0)
return 1;
return 0;

039-045: Waits until all threads finish executing

Parallelism using a Framework

Many frameworks exist to aid in parallelization, but the ones used in practical applications, such as in research labs and retail products, are limited. The most widely used frameworks are Message Passing Interface (MPI) for cluster servers, OpenMP for shared memory systems (SMP, NUMA), and the parallelization APIs in Boost C++. These frameworks require the user to specify the section as well as the method used for parallelization, but take care of tasks such as data transfer and execution of the threads, allowing the user to focus on the main core of the program.

List 1.3 shows an example usage of OpenMP, which is supported by most mainstream compilers, such as GCC, Intel C, Microsoft C. From the original sequential code, the programmer inserts a "#pragma" directive, which explicitly tells the compiler which sections to run in parallel. The compiler then takes care of the thread creation and the commands for thread execution.

List 1.3: OpenMP Example

#define N 100
#define TH_NUM 4
int main ()
int i;
int rootBuf[N];
/* Initialize array*/
rootBuf[i] = i;
/* Parallel process */
#pragma omp parallel for
for (i = 0; i < N; i++) {
rootBuf[i] = rootBuf[i] + 1;

When compiling the above code, the OpenMP options must be specified. GCC (Linux) requires "-fopenmp", intel C (Linux) requires "-openmp", and Microsoft Visual C++ requires /OpenMP. The code is explained below.


003: Include file required to use OpenMP

#define N 100  

004: Size of the array, and the number of times to run the loop. In general, this number should be somewhat large to benefit from parallelism.


012: Specify the number of threads to be used. The argument must be an integer.

#pragma omp parallel for 

020: Breaks up the for loops that follows this directive, into the number of threads specified in 012.

This example shows how much simpler the programming becomes if we use the OpenMP framework. Compare with List 2 that uses pthreads.

Automatic parallelization compiler

Compilers exist that examines for-loops to automatically decide sections that can be run in parallel, as well as how many threads to use.

For example, Intel C/C++ compiler does this when the option is sent.

(On Linux)

> icc -parallel -par-report3 -par-threshold0 -03 o parallel_test parallel_test.c


> Icc /Qparallel /Qpar-report3 /Q-par-threshold0 -03 o parallel_test parallel_test.c

The explanations for the options are discussed below.

  • -parallel: Enables automatic parallelization
  • -par-report3: Reports which section of the code was parallelized. There are 3 report levels, which can be specified in the form -par-report[n]
  • -par-threshold0: Sets the threshold to decide whether some loops are parallelized. In order to benefit from parallelization, enough computations must be performed within each loop to hide the overhead from process/thread creation. This is specified in the form -par-threshold[n]. The value for n must be between 0 ~ 100, with higher number implying higher number of computations. When this value is 0, all sections that can be parallelized become parallelized. The default value is 75.

At a glance, the automatic parallelization compiler seems to be the best solution since it does not require the user to do anything, but in reality, as the code becomes more complex, the compiler has difficulty finding what can be parallelized, making the performance suffer. As of August 2009, no existing compiler (at least no commercial) can auto-generate parallel code for hybrid systems such as the accelerator.


Fixstars  Solutions

Social Links