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
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
|
This is a partial list of the known problems in MPICH.
Current as of Release 1.1.2
General:
========
1. Cancel of nonblocking sends doesn't work
2. When using strict ANSI/ISO C compilation, the use of the "long long" type
generates an error. This is tricky; "long long" is needed by ROMIO on some
systems to get 8-byte integers for file offsets. Basically, if you want large
files, you may need to accept "long long".
3. The Fortran 90 compile and Fortran compilers may be passed the same
shared-library flags (e.g., -PIC). Some Fortran 90 compilers may not
recognize the same flags, and generate an "unrecognized argument" warning.
4. Totalview access to the message queues may not work. We've done some
preliminary testing but due to bureaucratic snafus haven't been able to fully
test this.
5. Debugger switches to mpirun (e.g., -dbx, -gdb) may not work. This is
caused by differences in the command line syntax for the debuggers; even for a
single vendor, the command line syntax for the debugger may change from month
to month. We plan to offer a more generalized configuration method for this
in the future.
6. MPICH is not thread-safe. A thread-safe version is in our plans.
7. The choice of default struct alignment is fixed and may not match your
compiler. The best fix is to use an MPI_UB when using MPI_Type_struct to
define the end of the structure. A subsequent version of MPICH will provide
more flexibility on structure alignment (requires changes to MPI_Type_struct).
8. Shared libraries for MPI. We're still working on this; the lack of a
common method for compiling/creating/using them is a burden. Still, there are
some steps in this direction; see mpich/util/makesharedlib. If you can help
us add to systems for which we can generate shared libraries, please send mail
to mpi-bugs@mcs.anl.gov .
9. If your Fortran and Fortran90 compilers use different name mappings (e.g.,
one uses a single trailing underscore and the other uses two), you won't be
able to use one of them. Systems that support weak symbols may be fixed in a
later release as long as other interface issues (e.g., size of MPI_Fint) are
the same.
10. Performance of collective routines is inadequate. Many of these are
simple functional implementations. If you would like to contribute better
versions, please contact us at mpi-bugs@mcs.anl.gov . One project to
provide better collective operations is known to us: MagPIe at
http://www.cs.vu.nl/albatross/ .
11. The installation process doesn't install the HTML versions of the
documentation.
12. When configuring with the -f95nag switch or the -fc=f95 switch, the
variable F77 is assigned the value of f95 and is substituted in the mpif77
script. Thus, the mpif77 script is used to run f95 programs. This can be
very misleading.
13. Systems with a Unix System V heritage may show performance problems when
using the ch_p4 device. This is due to the P4 code working around limitations
in the SysV/R2 interface. Later versions don't have the limitation that P4 is
catering to. We have added some code to bypass this old code (the P4
configure tests to make sure that the problem has been fixed). The option
to use with the MPICH configure is -p4_opts=--enable-nonblockread . We have
*not* tested this option much yet, and if something goes wrong, you may see
error messages from MPICH/P4 about "Unexpected EOF on socket".
MPE:
====
1. MPE Logging doesn't currently record communicator information. This can
cause upshot/jumpshot/nupshot to become confused about message arrows.
2. MPE Logging uses the MPICH definitions for MPI_PROC_NULL, MPI_ANY_TAG, and
MPI_ANY_SOURCE. This is fine for MPICH, but can cause problems when using MPE
logging with other MPI implementations.
3. The mechanism that upshot/nupshot/jumpshot uses to match sends and receives
can be fooled by multiple messages from the same source and with the same tag,
particularly when nonblocking sends and receives are used.
JUMPSHOT:
========
1. There is no man pages for the java routines.
ROMIO:
======
1. There are problems with NFS (they are not in ROMIO). See the README.
2. Error messages in ROMIO do not use the MPI Error_string service yet.
3. ROMIO Fortran tests may use getarg; some systems require pxfgetarg.
Device Specific:
===============
ch_p4 device:
1. MPI_Abort may not abort all processes
2. In certain cases, established connections can be lost by the TCP layer
(this is under the MPICH layer). MPICH depends on TCP to be reliable; when
the TCP implementation decides that a connection has failed, so does MPICH.
It turns out that the algorithm that some TCP implementations uses to decide
that a connection has failed can confuse network congestion and local node CPU
load with a failed connection. To fix this, we will either need to use UDP
instead of TCP or add our own reliability layer on top of TCP.
This problem affects more than MPICH; we have xterms and Emacs sessions fail
because of this problem with TCP.
3: Heterogeneous systems containing LINUX on different architectures does not
work. The p4 layer that is used for workstation networks predates portable
operating systems. This will be fixed when the p4 layer is replaced.
ch_meiko:
1. This device is no longer supported. If someone would like to send us a
patch that would make it work, we'll make it available as part of the regular
release.
ch_lfshmem:
1. The "lock free shared memory" device requires very specific memory ordering
or operations; this usually requires either adding assembly language
instructions (e.g., a write sync) or special compiler options. You can't just
configure/make this device, alas.
System Specific:
================
Cray J90:
1. The Fortran interfaces may not compile because of MPIR_FROM_FLOG. This
is a macro that converts Fortran logical values to C booleans. The fix is to
put the argument to MPIR_FROM_FLOG into a temporary int variable and pass that
to MPIR_FROM_FLOG.
LINUX:
1. We currently don't distinguish between LINUX on an Alpha and LINUX on an
X86, either as an architecture (tarch) or within ch_p4 (see above under
ch_p4). This makes it impossible to have a single directory contain LINUX
builds for several different architectures for any device (e.g., ch_shmem).
2. System include files in 2.1 before 2.1.128 have errors than can keep MPICH
from compiling. Update your version of LINUX to fix these problems.
FreeBSD:
1. ch_shmem doesn't work. One one system, the semget call generates a SIGSYS
interrupt. This appears to be an error in FreeBSD, as semget should not
generate this signal (EACCESS, EINVAL, or ENOSPC should be used if the semget
calls fails in this circumstance).
Solaris86:
1. Some users have had trouble compiling with pgcc (in sys/signal.h); using
gcc fixed the problem.
|
