You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This MCA framework makes it possible to integrate new threading frameworks with the OpenMPI runtime.
5
+
6
+
BACKGROUND
7
+
----------
8
+
There has been a lot of interest in integrating alternative threading models, in particular lightweight threading models with MPI implementations. Open MPI's modular component architecture seems like an ideal architecture for this sort of integration (in fact it was, Open MPI used to support Solaris and Windows threads).
9
+
10
+
Recently there has been interest in integrating MPI with lightweight tasking layers, which led to work reviving and modernizing the old modular threading code but with an emphasis on integrating lightweight threading models.
11
+
12
+
SELECTING A THREADING MODEL
13
+
---------------------------
14
+
15
+
The threading model is chosen via the configure option --with-threads=<threading_model>. This will choose a compile time threading model as well as compiling the relevant MCA.
16
+
17
+
IMPLEMENTATION
18
+
--------------
19
+
The MCA for threading libraries is implemented in two places, once as a set of .h files in mca/threads/<threading_model>/threads_<threading_model>_{threads,mutex,tsd}.h which are defined inline to the main thread implementation and also as an mca component that is loaded at runtime.
20
+
21
+
For performance reasons, in particular synchonization overhead, it is not possible to implement a threading model as a traditional MCA. This means --at least in the short term-- that threading models are chosen at compile time rather than runtime options, using mechanisms similar to OpenMPI's libevent integration.
22
+
23
+
The .h files are meant to be run on the fast path containing inline synchonization functions (threads_<threading_model>_mutex.h, thread local storage (threads_<threading_model>_tsd.h) and the opal_thread structure (threads_<threading_model>_thread.h).
24
+
25
+
The rest of the threading implementation follows the normal MCA model:
26
+
27
+
* threads_<threading_model>_component.c describes the version of the module and specifies the module open behavior (the threading model initialization goes here).
28
+
29
+
* threads_<threading_model>_condition.c defines an instance of opal_condition_t which is used by condition.h to define Open MPI specific condition variables.
30
+
31
+
* threads_<threading_model>_event.c defines an interface to Open MPI's libevent hooks. It allows the threading module to use threading model specific memory allocation and synchronization structures with Open MPI's libevent integration.
32
+
33
+
* threads_<threading_model>_module.c defines the interface to opal's thread handle. It provides ways of comparing threads, getting the value of a thread via its handle and the implementation of thread local storage.
34
+
35
+
* threads_<threading_model>_mutex.c provides a slow path interface to creating and destroying mutices dynamically via mca allocation. They can also be defined statically using the .h fast path interface.
36
+
37
+
* threads_<threading_model>_wait_sync.c provides condition variable like waiting capability that ensures MPI progress while it waits.
38
+
39
+
TODO
40
+
----
41
+
Libevent integration with lightweight threading models is a work in progress. The current Open MPI libevent library assumes preemption and does not yield by default. Lightweight threading libraries typically require tasks to be cooperative and to voluntarily yield after some time.
42
+
43
+
Open MPI itself needs to be altered to use a common yielding model instead of usleep(3).
0 commit comments