GraphicsMagick has been transformed to use OpenMP for the 1.3 release series. OpenMP is a portable framework for accelerating CPU-bound and memory-bound operations using multiple threads. OpenMP originates in the super-computing world and has been available in one form or another since the late '90s.
Since GCC 4.2 has introduced excellent OpenMP support via GOMP, OpenMP has become available to the masses. Recently, Clang has also implemented good OpenMP support. Microsoft Visual Studio Professional 2005 and later support OpenMP so Windows users can benefit as well. Any multi-CPU and/or multi-core system is potentially a good candidate for use with OpenMP. Modern multi-core chipsets from AMD, Intel, IBM, Oracle, and ARM perform very well with OpenMP.
Most image processing routines are comprised of loops which iterate through the image pixels, image rows, or image regions. These loops are accelerated using OpenMP by executing portions of the total loops in different threads, and therefore on a different processor core. CPU-bound algorithms benefit most from OpenMP, but memory-bound algorithms may also benefit as well since the memory is accessed by different CPU cores, and sometimes the CPUs have their own path to memory. For example, the AMD Opteron is a NUMA (Non-Uniform Memory Architecture) design such that multi-CPU systems split the system memory across CPUs so each CPU adds more memory bandwidth as well. Server-class CPUs offer more independent memory channels than desktop CPUs do.
For severely CPU-bound algorithms, it is not uncommon to see a linear speed-up (within the constraints of Amdahl's law) due to the number of cores. For example, a two core system executes the algorithm twice as fast, and a four core system executes the algorithm four times as fast. Memory-bound algorithms scale based on the memory bandwith available to the cores. For example, memory-bound algorithms scale up to almost 1.5X on my four core Opteron system due to its NUMA architecture. Some systems/CPUs are able to immediately context switch to another thread if the core would be blocked waiting for memory, allowing multiple memory accesses to be pending at once, and thereby improving throughput. For example, typical speedup of 20-32X (average 24X) has been observed on the Sun SPARC T2 CPU, which provides 8 cores, with 8 virtual CPUs per core (64 threads).
An approach used in GraphicsMagick is to recognize the various access patterns in the existing code, and re-write the algorithms (sometimes from scratch) to be based on a framework that we call "pixel iterators". With this approach, the computation is restricted to a small unit (a callback function) with very well defined properties, and no knowledge as to how it is executed or where the data comes from. This approach removes the loops from the code and puts the loops in the framework, which may be adjusted based on experience. The continuing strategy will be to recognize design patterns and build frameworks which support those patterns. Sometimes algorithms are special/exotic enough that it is much easier to instrument the code for OpenMP rather than to attempt to fit the algorithm into a framework.
Since OpenMP is based on multi-threading, multiple threads access the underlying pixel storage at once. The interface to this underlying storage is called the "pixel cache". The original pixel cache code (derived from ImageMagick) was thread safe only to the extent that it allowed one thread per image. This code has now been re-written so that multiple threads may safely and efficiently work on the pixels in one image. The re-write also makes the pixel cache thread safe if a multi-threaded application uses an OpenMP-fortified library.
GraphicsMagick provides its own built-in 'benchmark' driver utility which may be used to execute a multi-threaded benchmark of any other utility command.
Using the built-in 'benchmark' driver utility, the following is an example of per-core speed-up due to OpenMP on a four-core AMD Opteron system (with Firefox and other desktop software still running). The image is generated dynamically based on the 'granite' pattern and all the pixel quantum values have 30% gaussian noise added:
% gm benchmark -stepthreads 1 -duration 10 convert \ -size 2048x1080 pattern:granite -operator all Noise-Gaussian 30% null: Results: 1 threads 5 iter 11.34s user 11.340000s total 0.441 iter/s 0.441 iter/cpu 1.00 speedup 1.000 karp-flatt Results: 2 threads 9 iter 20.34s user 10.190000s total 0.883 iter/s 0.442 iter/cpu 2.00 speedup 0.000 karp-flatt Results: 3 threads 14 iter 31.72s user 10.600000s total 1.321 iter/s 0.441 iter/cpu 3.00 speedup 0.001 karp-flatt Results: 4 threads 18 iter 40.84s user 10.460000s total 1.721 iter/s 0.441 iter/cpu 3.90 speedup 0.008 karp-flatt
Note that the "iter/s cpu" value is a measure of the number of iterations given the amount of reported CPU time consumed. It is an effective measure of relative efficacy since its value should ideally not drop as iterations are added. The karp-flatt metric is another useful metric for evaluating thread-speedup efficiency. In the above example, the total speedup was about 3.9X with only a slight loss of CPU efficiency as threads are added.
Often it is noticed that the memory allocation functions (e.g. from the standard C library such as GNU libc) significantly hinder performance since they are designed or optimized for single-threaded programs, or prioritize returning memory to the system over speed. Memory allocators are usually designed and optimized for programs which perform thousands of small allocations, and if they make a large memory allocation, they retain that memory for a long time. GraphicsMagick performs large memory allocations for raster image storage interspersed with a limited number of smaller allocations for supportive data structures. This memory is released very quickly since GraphicsMagick is highly optimized and thus the time between allocation and deallocation can be very short. It has been observed that some memory allocators are much slower to allocate and deallocate large amounts of memory (e.g. a hundred megabytes) than alternative allocators, even in single-threaded programs. Under these conditions, the program can spend considerable time mysteriously "sleeping".
In order to help surmount problems with the default memory allocators, the configure script offers support for use of Google gperftools 'tcmalloc', Solaris mtmalloc, and Solaris umem libraries via the --with-tcmalloc, --with-mtmalloc, and --with-umem options, respectively. When the allocation functions are behaving badly, the memory allocation/deallocation performance does not scale as threads are added and thus additional threads spend more time sleeping (e.g. on a lock, or in munmap()) rather than doing more work. Performance improvements of a factor of two are not uncommon even before contending with the hugh CPU core/thread counts available on modern CPUs. Using more threads which are slowed by poorly-matched memory allocation functions is wasteful of memory, system resources, human patience, and electrical power.
Many modern CPUs support "Turbo" modes where the CPU clock rate is boosted if only a few cores are active. When a CPU provides a "Turbo" mode, this decreases the apparent speed-up compared to using one thread because the one thread was executed at a much higher clock rate. Likewise, when a CPU becomes very hot (due to being heavily used), it may decrease its clock rates overall to avoid burning up, and this may also decreases the actual speed-up when using many threads compared to using one thread. Many CPUs support "hyperthreads" or other mechanisms in which one physical core will support multiple light-weight threads, and if the core is efficiently used by one thread, then this will decrease the apparent per-thread speed-up but the peak speed-up will hopefully still be bounded by the number of physical cores.
In most cases, OpenMP does not speed-up loading an image from a file, or writing an image to a file. It is common for file decode and encode to take longer than processing the image. Using uncompressed formats is recommended with a fast I/O subsystem (or in-memory 'blobs' in order to obtain the greated speed-up from OpenMP.
It has been observed that sometimes it takes much longer to start and stop GraphicsMagick than it takes for it to run the requested algorithm. The slowness is due to inefficiencies of the libraries that GraphicsMagick is linked with (especially the ICU library that libxml2 is often linked with). If GraphicsMagick takes too long to perform trivial operations, then consider using the 'modules' build, and investigate the 'batch' utility which allows running many GraphicsMagick commands as a 'batch' script. If a 'modules' build is not feasible, then configuring GraphicsMagick to only support the specific formats actually needed can help with its execution time and improve opportunity for OpenMP speed-up.
According to the OpenMP specification, the OMP_NUM_THREADS evironment variable may be used to specify the number of threads available to the application. Typically this is set to the number of processor cores on the system but may be set lower to limit resource consumption or (in some cases) to improve execution efficiency. The GraphicsMagick commands also accept a -limit threads limit type option for specifying the maximum number of threads to use.
Copyright © GraphicsMagick Group 2002 - 2023