1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
|
This directory contains correctness tests and performance microbenchmarks for
GASNet. For convenience, the mpi directory also contains some comparable MPI
performance microbenchmarks, some contributed from outside sources.
See the top-level GASNet README for information on running all the tests using
the GASNet Makefile-based test harness.
See comments at the top of each file for test description, and run any test with
--help to see usage information for individual tests.
Coding standards for GASNet tests:
---------------------------------
All GASNet test code must adhere to the following coding standards:
* All of the general linguistic coding standards listed in the top-level GASNet README-devel
* Tests should only use public GASNet interfaces (ie prefixed with gasnet_ or gasnett_).
Tests must never invoke gasneti_, gasnetc_ or gasnete_ interfaces - this ensures test
portability across GASNet conduits and GASNet implementations. Any tests of GASNet
internal functionality belong in ../gasnet_diagnostic.c, where they will be invoked
by testinternal.
* test.h contains numerous tools which capture common idioms for GASNet test
code. Test authors should be familiar with the services provided by this
header and use them rather than re-implementing ad-hoc solutions.
Specifically:
- All tests must invoke test_init immediately after attach to register usage info
and init the framework. Failures in argument parsing should call test_usage().
- Tests needing segment space should use TEST_SEGSZ_REQUEST and TEST_SEG instead
of gasnet_getSegmentInfo, as the former handles SEGMENT_EVERYTHING correctly
- Barriers should be written as BARRIER(), unless there's a reason to directly
invoke the GASNet barrier functions (like testing barrier correctness)
- Threaded barriers should use PTHREAD_BARRIER(numthreads) or PTHREAD_LOCALBARRIER(numthreads)
- Non-fatal errors should issue a message using ERR("msg",...) (works just like printf)
- Fatal errors should be written as FATALERR("msg",...) (works just like gasnett_fatalerror)
- Informational messages should use MSG("msg",...) instead of printf,
or MSG0("msg",...) if the message is collective and should only be issued by node 0
- Allocation should use test_malloc, test_calloc, test_free
- Calls to the core API should be surrounded with GASNET_Safe()
- Pthread creation should be done using test_createandjoin_pthreads
- Timing should be done with TIME() or gasnett_ticks_now() (latter provides ns granularity)
- Tests needing random numbers should use TEST_RAND and friends, and should
include an optional seed argument that is passed to TEST_SRAND
- Tests with many optional parts should use the TEST_SECTION* interface
- Tests needing a computational delay should use test_delay
In all cases, see existing tests for examples of correct usage.
* All else being equal, when writing new test code it's usually preferable to
expand an existing test with related functionality rather than creating a
whole new test, unless one is testing a truly novel aspect of functionality or
performance. This helps to reduce build time and nightly test cost for the tests.
* New tests should be added to tests/Makefile.in (with appropriate arguments, if necessary)
and also to tests/harness/gasnet-tests/harness.conf
|
