README.md 7.11 KB
Newer Older
1
# Margo
2

Philip Carns's avatar
Philip Carns committed
3 4 5 6 7 8 9 10 11 12 13 14
Margo provides Argobots-aware wrappers to common Mercury library functions.
It simplifies service development by expressing Mercury operations as
conventional blocking functions so that the caller does not need to manage
progress loops or callback functions.

Internally, Margo suspends callers after issuing a Mercury operation, and
automatically resumes them when the operation completes.  This allows
other concurrent user-level threads to make progress while Mercury
operations are in flight without consuming operating system threads.
The goal of this design is to combine the performance advantages of
Mercury's native event-driven execution model with the progamming
simplicity of a multi-threaded execution model.
15 16 17

See the following for more details about Mercury and Argobots: 

18
* https://mercury-hpc.github.io/
19
* https://collab.mcs.anl.gov/display/ARGOBOTS/Argobots+Home
20

Philip Carns's avatar
Philip Carns committed
21 22 23
A companion library called abt-io provides similar wrappers for POSIX I/O
functions: https://xgitlab.cels.anl.gov/sds/abt-io

24
Note that Margo should be compatible with any Mercury transport (NA plugin).  The documentation assumes the use of the NA SM (shared memory) plugin that is built into Mercury for simplicity.  This plugin is only valid for communication between processes on a single node.  See [Using Margo with other Mercury NA plugins](##using-margo-with-other-mercury-na-plugins) for information on other configuration options.
25

26 27 28
##  Dependencies

* mercury  (git clone --recurse-submodules https://github.com/mercury-hpc/mercury.git)
29
* argobots (git clone https://github.com/pmodels/argobots.git)
30
* libev (e.g libev-dev package on Ubuntu or Debian)
31
* (optional) abt-snoozer (git clone https://xgitlab.cels.anl.gov/sds/abt-snoozer)
32

33 34
### Recommended Mercury build options

35 36
* Mercury must be compiled with -DMERCURY_USE_BOOST_PP:BOOL=ON to enable the
  Boost preprocessor macros for encoding.
37 38 39
* -DMERCURY_USE_CHECKSUMS:BOOL=OFF disables automatic checksumming of all
  Mercury RPC messages.  This reduces latency by removing a layer of
  integrity checking on communication.
40 41 42
* Mercury should be compiled with -DMERCURY_USE_SELF_FORWARD:BOOL=ON in order to enable
  fast execution path for cases in which a Mercury service is linked into the same
  executable as the client
43

44 45 46 47 48
Example Mercury compilation:

```
mkdir build
cd build
49
cmake -DMERCURY_USE_SELF_FORWARD:BOOL=ON -DMERCURY_USE_CHECKSUMS:BOOL=OFF \
50 51 52 53 54
 -DBUILD_TESTING:BOOL=ON -DMERCURY_USE_BOOST_PP:BOOL=ON \
 -DCMAKE_INSTALL_PREFIX=/home/pcarns/working/install \
 -DBUILD_SHARED_LIBS:BOOL=ON -DCMAKE_BUILD_TYPE:STRING=Debug ../
```

55 56 57 58 59 60 61
## Building

Example configuration:

    ../configure --prefix=/home/pcarns/working/install \
        PKG_CONFIG_PATH=/home/pcarns/working/install/lib/pkgconfig \
        CFLAGS="-g -Wall"
Philip Carns's avatar
Philip Carns committed
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
## Running examples

The examples subdirectory contains:

* margo-example-client.c: an example client
* margo-example-server.c: an example server
* my-rpc.[ch]: an example RPC definition

The following example shows how to execute them.  Note that when the server starts it will display the address that the client can use to connect to it.


```
$ examples/margo-example-server na+sm://
# accepting RPCs on address "na+sm://13367/0"
Got RPC request with input_val: 0
Got RPC request with input_val: 1
Got RPC request with input_val: 2
Got RPC request with input_val: 3
Got RPC request to shutdown

$ examples/margo-example-client na+sm://13367/0
ULT [0] running.
ULT [1] running.
ULT [2] running.
ULT [3] running.
Got response ret: 0
ULT [0] done.
Got response ret: 0
ULT [1] done.
Got response ret: 0
ULT [2] done.
Got response ret: 0
ULT [3] done.
```

The client will issue 4 concurrent RPCs to the server and wait for them to
complete.

## Running tests

`make check`

## Using Margo with the other NA plugins

You can use either the CCI NA plugin or BMI NA plugin to use either the CCI or BMI library for remote communication.  See the [Mercury documentation](http://mercury-hpc.github.io/documentation/) for details and status.

### CCI
110

111 112
Add the -DNA_USE_CCI:BOOL=ON option to the Mercury configuration.

113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135
You must set the CCI_CONFIG environment variable to point to a valid CCI
configuration file.  You can use the following example and un-comment the
appropriate section for the transport that you wish to use.  Note that there
is no need to specify a port; Mercury will dictate a port for CCI to use if
needed.

```
[mercury]
# use this example for TCP
transport = tcp
interface = lo  # switch this to eth0 or an external hostname for non-localhost use

## use this example instead for shared memory
# transport = sm

## use this example instead for InfiniBand
# transport = verbs
# interface = ib0
```

You must then use addresses appropriate for your transport at run time when
executing Margo examples.  Examples for server "listening" addresses:

136 137 138
* cci+tcp://3344 # for TCP/IP, listening on port 3344
* cci+verbs://3344 # for InfiniBand, listening on port 3344
* cci+sm://1/1 # for shared memory, listening on CCI SM address 1/1
139 140 141

Examples for clients to specify to attach to the above:

142 143
* cci+tcp://localhost:3344 # for TCP/IP, assuming localhost use
* cci+verbs://192.168.1.78:3344 # for InfiniBand, note that you *must* use IP
144
  address rather than hostname
145
* cci+sm:///tmp/cci/sm/`hostname`/1/1 # note that this is a full path to local
146 147 148
  connection information.  The last portion of the path should match the
  address specified above

149
### BMI
150

151 152 153 154
Add the -DNA_USE_BMI:BOOL=ON option to the Mercury configuration.  You may
also need to specify
-DBMI_INCLUDE_DIR:PATH=/home/pcarns/working/install/include and -DBMI_LIBRARY:FILEPATH=/home/pcarns/working/install/lib/libbmi.a (adjusting the paths as appropriate for your system).

155
We do not recommend using any BMI methods besides TCP.  It's usage is very similar to the CCI/TCP examples above, except that "bmi+" should be substituted for "cci+".
Philip Carns's avatar
Philip Carns committed
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

## Design details

![Margo architecture](doc/fig/margo-diagram.png)

Margo provides Argobots-aware wrappers to common Mercury library functions
like HG_Forward(), HG_Addr_lookup(), and HG_Bulk_transfer().  The wrappers
have the same arguments as their native Mercury counterparts except that no
callback function is specified.  Each function blocks until the operation 
is complete.  The above diagram illustrates a typical control flow.

Margo launches a long-running user-level thread internally to drive
progress on Mercury and execute Mercury callback functions (labeled
```__margo_progress()``` above).  This thread can be assigned to a
dedicated Argobots execution stream (i.e., an operating system thread)
to drive network progress with a dedicated core.  Otherwise it will be
automatically scheduled when the caller's execution stream is blocked
waiting for network events as shown in the above diagram.

Argobots eventual constructs are used to suspend and resume user-level
threads while Mercury operations are in flight.

Margo allows several different threading/multicore configurations:
* The progress loop can run on a dedicated operating system thread or not
* Multiple Margo instances (and thus progress loops) can be 
  executed on different operating system threads
* (for servers) a single Margo instance can launch RPC handlers
  on different operating system threads