forked from hpc/Spindle
-
Notifications
You must be signed in to change notification settings - Fork 0
/
INSTALL
185 lines (148 loc) · 8.87 KB
/
INSTALL
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
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
=============================================================================
== SPINDLE INSTALLATION
=============================================================================
Wait! Were you about to type './configure && make && make install'?
Unfortunately, it's not going to be that easy. There's likely several
options you'll have to pass to Spindle's configure line before it'll
work with your cluster. In short, you may have to provide some subset
of the following to configure:
1) Specify a security model
2) Specify a compiler to use for both front-end and back-end processes
3) Specify a start-up mechanism that works on your cluster
4) Specify a back-end local storage location
5) Optionally specify any misc tuning options
You can run 'configure --help' to see all of Spindle's configuration
options, though most of these can be left as defaults. Here's some
details about the above options.
1) Specify a security model
Spindle creates a network of daemons, one running on each node, and
connected via TCP/IP connections. Any file's contents can be
requested through this network, and executable code can be sent along
the network to be run on other processes. If an external process
makes an TCP/IP connection into the network on the appropriate port,
they'll essentially be able to get complete control over someone's
account.
Spindle thus has several security models that authenticate connections
into its network. Each of these has trade offs or requirements. As of
this writing there are four configure options for setting the security
mode. If multiple options are provided then Spindle will compile in
support for each option, and a user can select the security model at
runtime:
--enable-sec-munge - This option tells Spindle to use munge for its
security model. Munge is an external package
(https://code.google.com/p/munge) that runs as a daemon on each
node in a cluster. It's fast, scalable and relatively easy to
install (though it requires a root install) authentication
mechanism. If Munge is available, it's Spindle's first choice for
a security mechanism. You can point Spindle at a munge
installation directory with the --with-munge-dir configure option.
--enable-sec-keydir=DIR - This option tells Spindle to use gcrypt to
authenticate connections. Spindle will share a key need by gcrypt
by writing it into a permission-restricted file in DIR. DIR will
need to be a shared file system mounted across your cluster. You
can specify environment variables in DIR, which (when properly
escaped) will be evaluated at Spindle's runtime. So if $HOME is
globally mounted, you could do
'--enable-sec-keydir=\$HOME/.spindle'. Note that this option will
cause some extra runtime overhead, as a daemon on each node in the
cluster will need to read the keyfile.
--enable-sec-launchmon - This option tells Spindle to use gcrypt to
authenticate connections. Spindle will share a key by
broadcasting it with LaunchMON's communication infrastructure.
Note that this makes the key only as secure as LaunchMON's
communications. The (currently pending) LaunchMON 1.0.0 release
should have its own security infrastrucure.
--enable-sec-none - This option tells Spindle not to authenticate
connections into its network. This should only be used in secure
and broadly-trusted environments. This option is never selected
by default, and must be explicitly passed on the command line.
2) Specify a compiler to use for both front-end and back-end processes
Spindle has multiple components that run on different parts of a
cluster. Spindle is expected to be launched from a front-end node,
where MPI jobs are launched. On back-end, where MPI jobs run, it also
expects to run a daemon and inject a library into applications.
If your cluster has a single compiler that can build libraries and
executables for the front-end and back-end nodes, then you don't have
to do much here. Make sure configure uses that compiler (set the
standard CC/CXX variables if necessary) and you're good to go.
If you need to use different compilers to build code for the front-end
and back-ends, then you need to point Spindle at those compilers. The
standard CC/CXX environment variables will be used to build code for
the front-end, while the BE_CC/BE_CXX environment variables are
expected to point at compilers that build code for the back-ends. For
example, if gcc/g++ build code for the front-ends, and icc/icpc build
code for the backends, then you could run configure with something
like:
configure CC=gcc CXX=g++ BE_CC=icc BE_CXX=icpc
Note that Spindle doesn't use MPI, so you may not want to pass
mpi-wrapper compilers to Spindle's configure.
There are other variables for controlling Spindle's compiler usage,
run 'configure --help' to see them all.
3) Specify a start-up mechanism that works on your cluster
Spindle prefers to use LaunchMON
(http://sourceforge.net/projects/launchmon) to start its daemons, but
that won't work on all clusters. LaunchMON is normally used for
starting debuggers on clusters, but Spindle has different requirements
than debuggers. One way to tell if Spindle/LaunchMON will work on
your system is to create an MPI job under your favorite debugger on
your cluster. If the debugger give you control of your processes
while they're in MPI_Init, then Spindle/LaunchMON won't work on your
cluster* (because Spindle needs to get control before libraries are
loaded, which happens before MPI_Init). If the debugger gives you
control at the first instruction in the process, then
Spindle/LaunchMON should also work.
* = OpenMPI is an exception to this rule. Normally when running with
OpenMPI a debugger you'll get control in MPI_Init, but Spindle knows
some tricks to make LaunchMON work with OpenMPI.
If LaunchMON is going to work on your cluster, then install it and
point configure at LaunchMON with the --with-launchmon configure
option.
If LaunchMON isn't going to work, then you'll need to run Spindle in
"hostbin" mode. This may involve writing a script that plugs into
Spindle. Without LaunchMON, Spindle needs an alternative mechanism
for obtaining the list of back-end hostnames that are associated with
an MPI job, which can be different on each cluster. The
--with-hostbin=EXE configure option allows external services to
provide this information. When starting a job Spindle will execute
the script/executable specified by EXE. The MPI launcher's stdout and
stderr are piped into EXE's stdin, and the PID of the job launcher
process is passed to EXE on the command line. EXE should print the
list of hosts that are running job processes, separated by newlines,
to stdout and then exit with a zero return code.
4) Specify a back-end local storage location
Spindle is transmiting library files to each back-end, and it needs a
location to store those libraries. That location does not need to be
shared across the cluster, and it should provide fast, scalable access
to files. A ramdisk or local SSD would be ideal.
The --with-localstorage=DIR option should be used to specify this
location. The DIR parameter can be an escaped environment variable,
which will be evaluated on the back-end nodes. The default value for
this option is --with-localstorage=\$TMPDIR.
5) Optionally specify any misc tuning options
There are a few other non-standard configure options you may care
about:
--with-python-prefix=DIRS - Spindle can provide a better
quality-of-service on Python programs if it knows the prefix where
Python was installed. This configure option provides a
colon-separated list of directories where Spindle may find Python
installations. The directories in DIRS are treated as prefixes,
and any file operation in their subdirectories are assumed to be
on Python files. This directory list should not contain any
directories where the application will make writes (so it would be
a bad idea to add '/' to this list).
--with-testrm=ARG - Spindle has a testsuite, which can be invoked by
running $SPINDLE_BUILDDIR/testsuite/runTests. The testsuite needs
to know how to launch MPI jobs on your system. When given this
option, Spindle's testsuite will look for a script at
$SPINDLE_SRCDIR/testsuite/run_driver_$ARG that invokes MPI jobs.
See the testsuite/run_driver_template for an example script. As
of this writing, Spindle ships with with run_driver_flux,
run_driver_openmpi, and run_driver_slurm.
--with-usage-logging=FILE - Spindle can write an entry into FILE
every time it's invoked. This entry records the user who ran
spindle, the time it was run, and the Spindle version.
--with-default-numports=NUM
--with-default-port=NUM - These options specify a range of TCP/IP
ports that Spindle will use for communicating between its daemons.
It's usually okay to leave these at their default values (port
range 21940-21964), unless there happens to be a conflict.