The Habanero-C (HC) language under development in the Habanero project at Rice University builds on past work on Habanero-Java , which in turn was derived from X10 v1.5. HC serves as a research testbed for new compiler and runtime software technologies for extreme scale systems for homogeneous and heterogeneous processors. Like HJ, HC implements an execution model for multicore processors based on four orthogonal dimensions for portable parallelism:
1. Lightweight dynamic task creation and termination using async, finish, data-driven futures.
2. Collective and point-to-point synchronization using phasers
3. Mutual exclusion and isolation using isolated (not yet supported)
4. Locality control using hierarchical place trees
Unlike HJ which needs a JVM to run, Habanero-C is designed to be mapped onto hardware platforms with lightweight system software stacks, such as the Customizable Heterogeneous Platform (CHP) being developed in the NSF Expeditions Center for Domain-Specific Computing (CDSC) which includes CPUs, GPUs, and FPGAs. The C foundation also makes it easier to integrate HC with communication middleware for cluster systems, such as MPI and GASNet.
The Habanero-C compiler is written in C++ and is built on top of the ROSE compiler infrastructure, which is also used in the DARPA-funded PACE project under way at Rice University. The bulk of the Habanero-C runtime has been written from scratch in portable ANSI C. However, a few library routines for low-level synchronization and atomic operations are written in assembly language for the target platform. To date, the Habanero-C runtime has been ported and tested on Intel X86, Cyclops 64 and Intel SCC multicore platforms.
A short summary of the HC language is included below. Details on the underlying implementation technologies can be found in the Habanero publications web page. The HC implementation is still evolving at an early stage. If you would like to try out HC, please contact one of the following people: Zoran Budimlić, Yonghong Yan, Vincent Cave, or Vivek Sarkar.
- Habanero-C has two basic primitives for the task parallel programming model borrowed from X10: async and finish. The async statement, async <stmt>, causes the parent task to fork a new child task that executes <stmt>. Execution of the async statement returns immediately, i.e., the parent task can proceed to its following statement without waiting for the child task to complete. The finish statement, finish <stmt>, performs a join operation that causes the parent task to execute <stmt> and then wait until all the tasks created within <stmt> have terminated (including transitively spawned tasks). The Habanero-C runtime uses a work-stealing scheduler that supports work-first and help-first policies along with places for locality
- Habanero-C uses phasers for synchronization. Phasers are programming constructs that unify collective and point-to-point synchronization in task parallel programming. Phasers are designed for ease of use and safety, helping programmer productivity in task parallel programming and debugging. The use of phasers guarantees two safety properties: deadlock-freedom and phase-ordering. These properties, along with the generality of its use for dynamic parallelism, distinguish phasers from other synchronization constructs such as barriers, counting semaphores and X10 clocks.
- For locality, Habanero-C uses Hierarchical Place Trees (HPTs). HPTs abstract the underlying hardware using hierarchical trees, allowing the program to spawn tasks at places, which for example could be cores, groups of cores sharing cache, nodes, groups of nodes, or other devices such as GPUs or FPGAs. The work-stealing runtime takes advantage of the hardware hierarchy to preserve locality when executing tasks.
HC Language Summary
async [(place)] [IN (var1, var2, ...)] [OUT (var1, var2, ...)] [INOUT (var1, var2, ...)] [await (ddf1, ddf2, ...)] [phased [(ph1, ph2, ...)]] Stmt
- asynchronously start a new task to execute Stmt- a destination place can optionally be specified for where the task should execute. The place can be obtained from the runtime using HC runtime functions:
hc_get_place(NVGPU_PLACE | MEM_PLACE | FPGA_PLACE) --- returns the place of the execution device
hc_get_child_place() --- returns the child place in the HPT that is on the path from the current place to the leaf place of the currently running worker
- the variables that the acync uses have to be specified in the IN(variables read by the async), OUT(variables written by the async and INOUT(variables both read and written by the async) clauses.
- an await clause can optionally be specified, listing all the data-driven futures (DDF's) that the task should wait on before starting its execution.
- a phased clause can optionally be specified, registering the async on all the phasers specified in the list (ph1, ph2, ...), or on all the phasers of the parent (if the list is not specified).
- execute Stmt, but wait until all (transitively) spawned asyncs in Stmt's scope have terminated
DDF_CREATE() --- a library function that creates a DDF, and returns a pointer to a DDF_t type
DDF_GET(DDF_t * ddf) --- perform a get on the value of a DDF. If the ddf is empty, the runtime will exit with an assertion failure.
DDF_PUT(DDF_t * ddf, void * value) --- perform a put of the value into a DDF ddf. If ddf already has a value filled in, the runtime will exit with an assertion failure.
async await (ddf1, ddf2, ...) Stmt --- wait until all the DDF's in the list (ddf1, ddf2, ...) have their values filled in before asynchronously starting the execution of Stmt. Stmt should only perform the GET's on the DDF's specified in the list, doing otherwise may result in undefined behavior.
Hierarchical Place Trees (HPT's)
Hierachical place tree of the machine you want to execute your HC program on is specified in an .xml file. For example, the .xml file for a single Sugar computational node looks like this:
<!DOCTYPE HPT SYSTEM "hpt.dtd">
<HPT version="0.1" info="an HPT for SUGAR, 2 quad core2 Intel Xeon processors">
<place num="1" type="mem">
<place num="2" type="cache"> <!-- 2 sockets -->
<place num="2" type="cache"> <!-- 2 L2 cache per socket -->
<place num="2" type="cache"> <!-- 2 L1 cache per L2 -->
You specify the HPT of the machine when invoking your HC executable, for example:
./a.out -nproc 8 -hpt sugar.xml
Some HC functions that allow you to navigate the HPT:
short is_cpu_place(place_t * pl) --- is pl a CPU place?
short is_device_place(place_t * pl) --- is pl a device (GPU or FPGA) place?
short is_nvgpu_place(place_t * pl) --- is pl a NVIDIA GPU place?
place_t* hc_get_current_place() --- get the place where the task is currently executing
int hc_get_num_places(short type) --- get the number of places of the specified type (NVGPU_PLACE, MEM_PLACE or FPGA_PLACE)
void hc_get_places(place_t ** pls, short type) --- get an array of all the places of the specified type
place_t * hc_get_place(short type) --- get any place of the specified type
place_t * get_ancestor_place(hc_workerState * ws)
place_t * hc_get_child_place() --- get the child place on the path from the current place to the leaf place of the current worker
place_t * hc_get_parent_place() --- get the parent place of the current place in HPT
place_t ** hc_get_children_places(int * numChildren) --- get an array of all the child places of the current place
Current HC limitations
There are some limitations due to the current implementation of the HC programming model. These limitations are not inherent to the programming model, but rather are a result of some incompleteness in our compiler or runtime implementation.
- pointers to stack variables cannot be passed to async tasks. This can result in unpredictable behavior. Workaround: convert all the data meant to be shared among tasks to heap-allocated
- pointers to HC functions (functions that contain HC constructs or call other HC functions) are not allowed in HC
- 'const' modifiers are not allowed for function parameters or local variables in HC programs. Workaround: remove the 'const' modifiers. The program semantic will remain unchanged, as the only purpose of the 'const' modifiers is to enforce some additional compiler checking.
- number of tasks registered on a phaser cannot be bigger than the number of worker threads specified with the -nproc option when the HC-compiled executable is invoked. Otherwise, a deadlock can occur.