README.md 8.66 KB
Newer Older
1
# Margo
2

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

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

116 117 118
* 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
119 120 121

Examples for clients to specify to attach to the above:

122 123
* 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
124
  address rather than hostname
125
* cci+sm:///tmp/cci/sm/`hostname`/1/1 # note that this is a full path to local
126 127 128
  connection information.  The last portion of the path should match the
  address specified above

129
### BMI
130

131 132 133 134
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).

135
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+".
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

## 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

165 166 167 168 169
## V0.2 API changes

The following list provides details on new changes to the Margo API starting in
version 0.2:

170 171
* `margo_init()` is much more simplified and initializes Mercury and Argobots on behalf of
  the user (with `margo_finalize()` finalizing them in that case)
Shane Snyder's avatar
Shane Snyder committed
172
    * the prototype is `margo_init(addr_string, MARGO_CLIENT_MODE | MARGO_SERVER_MODE, use_progress_thread, num_rpc_handler_threads)`
173 174
    * `margo_init_pool()` is still available as an advanced initialize routine, where the
      user must initialize Mercury/Argobots and pass in an `HG_Context` and `ABT_pools` for
175 176 177
      RPC handlers and the progress loop
* Margo now has its own RPC registration functions that should be used for registering
  RPCs (as Margo is now attaching internal state to RPCs)
178
    * `MARGO_REGISTER` is basically equivalent to `MERCURY_REGISTER`, except it takes a
Shane Snyder's avatar
Shane Snyder committed
179
      `margo_instance_id` rather than a Mercury class
180 181 182
    * `MARGO_REGISTER_MPLEX` is mostly the same as above, but allows a user to specify
      an `ABT_pool` to use for a given RPC type
* relatedly, Margo users should now use `margo_register_data()` (rather than `HG_Register_data()`)
183
  for associating user data with an RPC type
184
    * like Mercury, there is a corresponding `margo_registered_data()` call to retrieve the user pointer
185
* the following Mercury-like functions are now defined within Margo, although it is
Shane Snyder's avatar
Shane Snyder committed
186
  still safe to just use the Mercury calls directly (most are just `#define`s to the
187
  corresponding Mercury call, anyway):
188 189 190 191 192 193 194 195
    * `margo_registered_disable_response()`
    * `margo_addr_free()`
    * `margo_addr_self()`
    * `margo_addr_dup()`
    * `margo_addr_to_string()`
    * `margo_create()`
    * `margo_destroy()`
        * Note that `margo_create()`/`margo_destroy()` are enhancing `HG_Create()`/`HG_Destroy()` with a
196
          cache of reusable handles, so they may be preferable to the Mercury calls
197 198 199 200
    * `margo_bulk_create()`
    * `margo_bulk_free()`
    * `margo_bulk_deserialize()`
* `margo_hg_handle_get_instance()` and `margo_hg_info_get_instance()` calls have been added for
Shane Snyder's avatar
Shane Snyder committed
201
  retrieving a `margo_instance_id` given a received handle or the HG info structure associated
202 203
  with the handle