forked from sandialabs/portals4
-
Notifications
You must be signed in to change notification settings - Fork 1
/
README
164 lines (124 loc) · 6.46 KB
/
README
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
Portals 4 Reference Implementation
----------------------------------
* About
* Building
Portals utilizes the Autoconf/Automake/Libtool build system. The
standard GNU configure script ane make system is used to build Portals
4. To build:
% autoreconf (only once, or when the build system has been updated)
% ./configure <options>
% make
% make check
% make install
The "make check" step is not strictly necessary, but is a good idea.
Options to configure include:
--prefix=<DIR> Install implementation in <DIR>
* Portals 4 Implementation Notes
There is currently one Portals 4 implementation, ib. The ib
implementation supports multiple transports, including InfiniBand,
UDP/Reliable UDP (experimental), and SHMEM.
- Infiniband: this transport has multiple options, including not actualling using IB:
* Infiniband, selected by default, can be disabled with --disable-transport-ib
* shared memory, not selected by default, can be enabled with
--enable-transport-shmem. It has the option of using KNEM for the
large transfer (used if --with-knem=xxxx is present), or to
use an internal slower shared memory protocol.
Both Infiniband and shmem can be used at the same time. In that
case, Portals will use infinband between nodes, and shmem
intra-node. If only shmem is used, then portals will not run
between nodes (obviously).
On top of that, 2 different environments can be generated:
* the "fat library" which contains all the portals API. Each
process has its own progress thread which leads to the use of
too many cores on a machine. This is the default.
* the Portals Progres Engine (PPE) and the light library. The
PPE is a daemon that regroups the progress threads (one by
default) for all the portals process on the host. The portals
application links to the "light library" which is only a pass
through between the application and the PPE. Very litlle
processing is done in the light library; most of it is done
inside the PPE. This environment is selected with
--enable-ib-ppe. The XPMEM driver must be present, and so it
is mutually exclusive with the shmem transport. The daemon
must be started prior to running a Portals4 application:
<PATH>/p4ppe
p4ppe has a set of options. Run with the --help optins to see
them.
Both the fat and light library are called the same, thus only one
type can exist on a machine at the same time; use LD_LIBRARY_PATH
and/or LD_PRELOAD to work around.
Pre-requisites
* libev-4.0.4 (or similar), available at
http://software.schmorp.de/pkg/libev.html
* The KNEM driver from
http://runtime.bordeaux.inria.fr/knem/. Will be used for
shared memory if --with-knem=<path> is set. Ensure that
/dev/knem is readable/writable by the user running the portals
software.
* The XPMEM driver from
https://code.google.com/p/xpmem/. Mandatory if the PPE is
selected, unused otherwise. The current driver has crashing
issues and doesn't compile on recent kernel versions.
* The ummunotify driver from
https://github.com/Portals4/ummunotify/
(required for correctness with the IB transport). Ensure that /dev/ummunotify is
readable/writable by the user running the portals software.
WARNING: Not using ummunotify may result in program incorrectness..
Build:
Note that the paths to mpi may not be necessary or may need to
be changed depending on your linux distribution.
* minimal, no shared memory:
./configure
* optimized, with shared memory and KNEM support:
./configure --enable-transport-shmem --with-knem=/opt/knem --enable-fast
* PPE, no IB transport (so local node only):
./configure --enable-ib-ppe --disable-transport-ib
Then type "make".
Test:
* Type "make check" in the main directory. This will run a
series of checks on the local node, with either 1 or 2 ranks.
Configure will attempt to figure out the proper way to launch
parallel processes. The TEST_RUNNER environment variable can
be used to set the proper test runner (such as
TEST_RUNNER='yod -np $(NPROCS)').
Environment variables:
* PTL_ENABLE_MEM=[0|1] will deactivate/activate the local memory
transport, if one is compiled in.
* PTL_DEBUG=1 will activate the tracing
* PTL_LOG_LEVEL=[0|1|2|3] will set the trace level.
* PTL_IFACE_NAME allows for the explicit naming of the network interface
for example ib0, eno1, etc.
* PTL_DISABLE_MEM_REG_CACHE=[0|1] deactivates/activates the IB memory
registration cache. Disabling it no longer requires ummunotify, and
the implementation does not keep a registered memory cache.
For instance:
PTL_LOG_LEVEL=3 PTL_DEBUG=1 yod -n 1 ./spam
Building an RPM:
* make dist --> this will create portals-4.0.tar.gz
* rpmbuild -ta portals-4.0.tar.gz
Limitations:
* A program may not open more than one NID in its lifetime.
Why Portals may not run properly:
* The user cannot pin enough memory. By default linux systems
allow about 32KB. Check with "ulimit -l". This is for IB
only. Edit /etc/security/limits.co onf and change/add the
following 2 lines (including the stars):
* hard memlock unlimited
* soft memlock unlimited
(This setting allow every user to pin large amounts of
memory. Replace the stars with a user name for a tighter
setting).
* IB is not up and running (check with "ibv_devinfo") or the
default IPoIB interface, ib0, doesn't have an IP address. All
interfaces should be on the same IP range.
* slow performances can be caused by cpuspeed. Disable it
("service cpuspeed stop").
* uneven performance between repeats can be caused by the
threads being on different cores. Pin them.
* When shmem is used, under the wrong circumstances, the files
/dev/shm/portals4-shmem-* are not removed. If a subsequent run
tries to re-use the same name, it will fail and the
application may exit.
* If you get a mr_lookup: Assertion `res == ((void *)0)' failed
error, you need to configure Portals 4 with --enable-zero-mrs
to enable zero length MR suport.