From 33e713be50d5d8c1dcf3ad577f02728ed0cdabd0 Mon Sep 17 00:00:00 2001
From: Matthias Maier
- In a similar vain as for build configurations, it is possible to restrict
+ In a similar vein as for build configurations, it is possible to restrict
tests to specific feature configurations, e.g.,
If a test should be run with MPI in parallel, the number of MPI
processes category/test.output
. In pseudo code:
category/test.[with_<string>(<=|>=|=|<|>)<on|off|version>.]*
- [mpirun=<N|all>.][threads=<N|all>.][expect=<y>.][binary.][<debug|release>.](output|run_only)
+ [mpirun=<N|all>.][threads=<N|all>.][expect=<y>.][exclusive.][<debug|release>.](output|run_only)
Normally, a test will be set up so that it runs twice, once in debug and
once in release configuration.
@@ -500,7 +499,7 @@ category/test.release.output
Restricting tests to feature configurations
category/test.with_umfpack=on.output, or
@@ -543,7 +542,7 @@ category/test.with_umfpack=on.with_zlib=on.output
-
It is quite typical for an MPI-enabled test to have multiple output
files for different numbers of MPI processes.
+
+ Similarly, the thread pool size of a test can be specified by using
+ Running tests with MPI
+ Running tests with MPI or a specific thread pool size
N
with which a program needs to be run for
@@ -553,6 +552,38 @@ category/test.mpirun=N.output
threads=N
, where N
is the number of
+ concurrent worker threads that should be initialized:
+
+category/test.threads=N.output
+
+ This is declaration is equivalent to setting the environment variable
+ DEAL_II_NUM_THREADS
, or calling
+ MultithreadInfo::set_thread_limit()
by hand.
+
+ In order to account for the increased computational workload of MPI
+ parallel code the testsuite will add a processing weight to
+ individual MPI tests equal to half of the number of MPI ranks that
+ ensures that the machine the testsuite runs on is at most
+ overcommitted by a factor of 2. Thread concurrency is not accounted
+ for.
+
+ This behavior is modified for performance tests. Here, the
+ processing weight is taken directly to be the product of the number
+ of MPI ranks times the number of threads.
+
+ Particularly sensitive timing tests that have to be run
+ exclusively without any other test running concurrently can be
+ annotated with the .exclusive
keyword:
+
+category/test.exclusive.output ++ This ensures that the test in question always runs "in serial" without + another test scheduled concurrently. diff --git a/doc/users/testsuite.html b/doc/users/testsuite.html index ccc40b6819..149e5a17ef 100644 --- a/doc/users/testsuite.html +++ b/doc/users/testsuite.html @@ -322,13 +322,12 @@ TEST_MPI_RANK_LIMIT dropped. The special value 0 enforces no limit. Defaults to 0. TEST_THREAD_LIMIT - - Specifies the maximal number of worker threads that can should be - used by the threading backend. If a test variant configures a larger - number of threads (via .threads=N. in the output file) than this limit - the test will be dropped. Note that individual tests might exceed this - limit by calling MultithreadInfo::set_thread_limit(), or by manually - creating additional threads. The special value 0 enforces no limit. - Defaults to 0. + - Specifies the maximal number of worker threads that can be used by the + threading backend. If a test variant configures a larger number of threads + (via .threads=N. in the output file) than this limit the test will be + dropped. Note that individual tests might exceed this limit by calling + MultithreadInfo::set_thread_limit(), or by manually creating additional + threads. The special value 0 enforces no limit. Defaults to 0.