Mapping onto Specific Hardware Architectures¶
By providing an accelerator independent interface for kernels, their execution and memory accesses at different hierarchy levels, alpaka allows the user to write accelerator independent code that does not neglect performance.
The mapping of the decomposition to the execution environment is handled by the back-ends provided by the alpaka library as well as user defined back-ends.
A computation that is described with a maximum of the parallelism available in the redundant hierarchical parallelism abstraction can not be mapped one to one to any existing hardware.
GPUs do not have vector registers for
Therefore, the element level is often omitted on CUDA accelerators.
CPUs in turn are not (currently) capable of running thousands of threads concurrently and do not have equivalently fast inter-thread synchronization and shared memory access as GPUs do.
A major point of the redundant hierarchical parallelism abstraction is to ignore specific unsupported levels and utilize only the ones supported on a specific accelerator. This allows a mapping to various current and future accelerators in a variety of ways enabling optimal usage of the underlying compute and memory capabilities.
The grid level is always mapped to the whole device being in consideration. The scheduler can always execute multiple kernel grids from multiple queues in parallel by statically or dynamically subdividing the available resources. However, this will only ever simplify the mapping due to less available processing units. Furthermore, being restricted to less resources automatically improves the locality of data due to spatial and temporal locality properties of the caching hierarchy.
There are multiple possible ways to map the alpaka abstraction to x86 CPUs. The following figure shows the compute and memory hierarchy of a dual-socket (package) node with dual-core CPUs and symmetric multithreading (Hyper-Threading). Through symmetric multithreading (Hyper-Threading) each core represents two processing units.
Mapping the thread level directly to the processing units is the most trivial part of the assignment of hierarchy levels to hardware units. However, the block and warp levels could be mapped to hardware components in different ways with varying advantages and disadvantages.
Even though a warp seems to be identical to a vector register, because both execute a single uniform instruction on multiple data elements, they are not the same. Warps can handle branches with divergent control flows of multiple threads. There is no equivalent hardware unit in a CPU supporting this. Therefore, the warp level can not be utilized on CPUs leading to a one-to-one mapping of threads to warps which does not violate the rules of the abstraction.
One Block Per Node¶
By combining all processing units (possibly Hyper-Threads) of all processors on a node into one block, the number of synchronizing and communicating threads can be enlarged. This high possible thread count would simplify the implementation of some types of algorithms but introduces performance issues on multi-core nodes. The shared memory between all cores on a node is the RAM. However, the RAM and the communication between the sockets is far too slow for fine-grained communication in the style of CUDA threads.
One Block Per Socket¶
If each processor on each socket would concurrently execute one block, the L3 cache would be used as the fast shared memory. Although this is much better then to use the RAM, there is still a problem. Regions of the global memory and especially from the shared memory that are accessed are automatically cached in the L1 and / or L2 caches of each core. Not only the elements which are directly accessed will be cached but always the whole cache line they lie in. Cache lines typically have a size of 64 Bytes on modern x86 architectures. This leads to, for example, eight double precision floating point numbers being cached at once even though only one value really is required. As long as these values are only read there is no problem. However, if one thread writes to a value that is also cached on other cores, all such cache copies have to be invalidated. This results in a lot of cache and bus traffic. Due to the hierarchical decomposition of the grid of threads reflecting the data elements, neighboring threads are always combined into a common block. By mapping a block to a socket, threads that are executed concurrently always have very close indices into the grid. Therefore, the elements that are read and written by the threads are always very close together within the memory and will most probably share a cache line. This property is exploited on CUDA GPUs, where memory accesses within a warp are combined into one large transaction. However, when multiple threads from multiple CPU cores write to different elements within a cache line, this advantage is reversed into its opposite. This pattern non-intuitively leads to heavy performance degradation and is called false-sharing.
One Block Per Core¶
The best compromise between a high number of threads per block and a fast communication between the threads is to map a block directly to a CPU core. Each processing unit (possibly a Hyper-Thread) executes one or more threads of our hierarchical abstraction while executing multiple elements locally either by processing them sequentially or in a vectorized fashion. This possible mapping of blocks, threads and elements to the compute and memory hierarchy of a dual-socket node with dual-core CPUs and symmetric multithreading is illustrated in the following figure. ![x86_cpu](x86/x86_cpu_mapping.png)
One Block Per Thread¶
If there is no symmetric multithreading or if it is desired, it is also possible to implement a mapping of one block with exactly one thread for each processing unit. This allows to completely remove the synchronization overhead for tasks where this is not required at all.
The mapping of threads to processing units is independent of the threading mechanism that is used. As long as the thread affinity to cores can be set correctly, OpenMP, pthread, std::thread or other libraries and APIs can be used interchangeably to implement various alpaka back-ends. They all have different advantages and disadvantages. Real operating system threads like pthread, std::thread and others have a high cost of thread creation and thread change because their default stack size amounts to multiple megabytes. OpenMP threads on the other hand are by default much more lightweight. However, they are arbitrarily limited by the runtime implementation in the maximum number of concurrent threads a machine supports. All of the previous methods have non-deterministic thread changes in common. Therefore it is not possible to decide the order in which threads within a block are processed, which could be a good optimization opportunity.
To allow blocks to contain more threads then the number of processing units each core provides, it is possible to simply start more threads then processing units are available. This is called oversubscription. Those threads can be bound to the correct cores and by relying on the operating system thread scheduler, they are preemptively multitasked while sharing a single cache and thereby avoiding false-sharing. However, this is not always beneficial because the cost of thread changes by the kernel-mode scheduler should not be underestimated.
Mapping the abstraction to GPUs supporting CUDA and HIP is straightforward because the hierarchy levels are identical up to the element level. So blocks of warps of threads will be mapped directly to their CUDA/HIP equivalent.
The element level is supported through an additional run-time variable containing the extent of elements per thread. This variable can be accessed by all threads and should optimally be placed in constant device memory for fast access.