In PHISH lingo, a "minnow" is a stand-alone application which makes calls to the PHISH library. Minnows are typically small programs which perform a single task, e.g. they parse a string searcing for keywords and store statistics about those keywords. But they can also be large programs which perform sophisticated computations and make only occasional calls to the PHISH library. In which case they should probably be called "sharks" or "whales" ...
An individual minnow is part of a "school" of one or more duplicate minnows. One or more schools form a PHISH "net(work)" which compute in a coordinated fashion to perform a calculation. Minnows communicate with each other to exchange data via calls to the PHISH library.
This doc page covers the following topics:
This is a list of minnows in the minnow directory of the PHISH distribution. Each has its own doc page. Some are written in C++ (*.cpp), some in Python (*.py), some in both. If provided in both languages, their operation is identical, with any exceptions noted in the minnow doc page:
These are also 3 special minnows which can wrap stand-alone non-PHISH applications which read from stdin and write to stdout, so that they can be used as minnows in a PHISH net and communicate with other minnows:
These are also two simple codes which can be compiled into stand-alone non-PHISH executables. They are provided as examples of applications that can be wrapped by the "wrap" minnows:
The easiest way to understand how a minnow works with the PHISH library, is to examine a few simple files in the minnow directory. Here we list the count.py minnow, which is written in Python. There is a also a count.cpp minnow, written in C++, which does the same thing. The purpose of this minnow is to count occurrences of strings that it receives as datums:
1 #!/usr/local/bin/python as path to Python if desired 2 3 import sys 4 import phish 5 6 def count(nvalues): 7 if nvalues != 1: phish.error("Count processes one-value datums") 8 type,str,tmp = phish.unpack() 9 if type != phish.STRING: 10 phish.error("Count processes string values") 11 if hash.has_key(str): hashstr = hashstr + 1 12 else: hashstr = 1 13 14 def dump(): 16 for key,value in hash.items(): 17 phish.pack_int(value) 18 phish.pack_string(key) 19 phish.send(0) 20 21 args = phish.init(sys.argv) 22 phish.input(0,count,dump,1) 23 phish.output(0) 24 phish.check() 25 26 if len(args) != 0: phish.error("Count syntax: count") 27 28 hash = 29 30 phish.loop() 31 phish.exit()
On line 4, the Python minnow imports the phish module, which is provided with the PHISH distribution. Instructions on how to use this module, which wraps the C-interface to the PHISH library, are given in this section of the documentation.
The main program begins on line 21. The call to the phish.init is typically the first line of a PHISH minnow. When the minnow is launched, extra PHISH library command-line arguments are added which describe how the minnow will communicate with other minnows. These are stripped off by the phish.init function, and the remaining minnow-specific arguments are returned as "args". The phish.input and phish.output functions setup the input and output ports used by the minnow. A port is a communication channel by which datums arrive from other minnows or can be sent to other minnows. The PHISH input script sets up these connections, but from the minnow's perspective, it simply receives datums on its input port(s) and writes datums to its output port(s). See the next section for more discussion of ports.
There should be one call to phish.input for each input port the minnow uses. And one call to phish.output for each output port it uses. The call to the phish.check function on line 24 insures that the minnow as written is compatible with the way it is used in the PHISH input script, i.e. that the necessary input and output ports have been defined with valid hook styles.
The phish.input call specifies a callback function that the PHISH library will invoke when a datum arrives on that input port. In this case, the count minnow defines a count() callback function which stores a received string in a hash table (Python dictionary) with an associated count of the number of times it has been received.
On line 28, an empty hash table is initialized, and then the phish.loop function is called. This gives control to the PHISH library, which will wait for datums to be received, invoking the appropriate callback function each time one arrives.
The call to phish.input also defines a callback to the dump() function which is invoked when input port 0 is closed. This occurs when upstream minnows send the requisite number of "done" messages to the port. The dump() function sends the contents of the hash table to output port 0, one datum at a time. Each datum contains a unique string and its count.
The phish.loop function returns after invoking dump() and when all input ports are closed. The count minnow then calls the phish.exit function which will close its output port(s), and send "done" messages to downstream minnows connected to those ports.
This code structure is typical of many minnows:
As discussed above, ports are input/output communication channels by which a minnow receives datums from an upstream minnow or sends datums to a downstream minnow.
The number of ports that can be configured by a minnow varies between PHISH library backends. The ZMQ version of the library supports an unlimited number of input and output ports, while the MPI version of the library supports up to MAXPORT number of input ports and MAXPORT number of output ports. MAXPORT is a hardwired value in src/phish-mpi.cpp which is set to 16. It can be changed if needed, but note that all minnows which use the MPI version of the library must be re-built since they must all use a consistent value of MAXPORT when run together in a PHISH net.
Note that a PHISH input script may connect a particular minnow to other minnows in a variety of ways. This applies to both the styles of hooks that are specified and the number of minnows on the other end of each hook. Thus it is possible for the user to specify hooks in the input script which the minnow does not support or even define. Similarly, the input script may cause other minnows to send datums to the minnow which it does not expect or is unable to interpret. This means a minnow should be coded to follow these rules:
Here are other flexible attributes of input and output ports to note, all enabled by the hook command in a PHISH input script:
All of these scenarios can be setup by appropriate use of hook commands in a PHISH input script.
An additional issue to consider is whether a communication channel can be saturated or drop datums. Imagine a PHISH net where one minnow sends datums at a high rate to a receiving minnow, which cannot process them as fast as they are sent. Over time, the receiving minnow is effectively a bottleneck in processing a stream of data. The PHISH library will not lose messages in this scenario, rather the overall processing pipeline naturally throttles itself to the rate of the bottlenecking minnow. This is handled by the underlying MPI or socket message passing protocols. ZMQ handles this naturally. In the case of MPI, the sending and receiving processes coordinate data exchanges. By default this is done via MPI_Send() and MPI_Recv() calls. If you get a run-time MPI error about dropping messages, then you should use occasionally use the "safe" mode of data exchange which can be enabled by the set safe command in a PHISH input script or "--set safe" command-line option. This will use MPI_Ssend() calls which enforce extra handshaking between the sending and receiving processes to avoid dropping messages.
PHISH minnows can be designed to process a finite or infinite stream of data. In the infininte case, the PHISH net of minnows is typically shut down by the user killing one or more of the processes. Note that the current ZMQ version of PHISH cannot guarantee that all processes will be shut-down cleanly. You may need to kill some of the procsses manually.
In the finite case, you typically want each minnow in the net to shut down cleanly.
The PHISH library sends special "done" messages when a minnow closes one of its output ports. This is triggered by a call to the phish_close function, which closes a single port, or the phish_exit function which closes all output ports. A "done message is sent to each receiving minnow of each input port connected to the corresponding output port. The receiving minnow counts these messages as they arrive. When it has received one "done" message from every minnow that connects to one of its input ports, it closes the input port and the library calls back to the minnow (if a callback function was defined by the phish_input function). When all its input ports have been closed it makes an additional callback to the minnow (if a callback function was defined by the phish_callback function).
This mechanism is often sufficient to trigger an orderly shutdown of an entire PHISH net by all its minnows, if the most upstream minnow initiates the process by closing its output ports via a call to phish_exit. Exceptions are when a school of minnows exchanges data in a "ring" style of commuication as setup by the hook ring command in a PHISH input script.
In case of the ring, if the first minnow in the ring invokes the phish_exit function, it will no longer be receiving datums when the last minnow in the ring attempts to send it a "done" message. In this case, the first minnow should instead invoke phish_close on the output port for the ring, then wait to receive its final "done" message before calling phish_exit.
Another exception is when minnows send datums to themselves in a looping fashion. In this case, you typically to include code in callback functions invoked when "done" messages are received to handle the shutdown logic. See the minnow/tri.py for an example of how this can be done.
Minnows are stand-alone programs which simply need to be linked with the PHISH library. New single-file minnows written in C or C++ can be added to the minnow directory of the PHISH distribution and built in the following manner; minnows written in Python do not need to be built.
The easiest way to build all of PHISH, including the PHISH minnows, is to use the cross-platform CMake build system. We recommend building PHISH with a separate build directory:
$ tar xzvf phish.tar.gz -C ~/src $ mkdir ~/build/phish $ cd ~/build/phish $ ccmake ~/src/phish-14sep12
Then, in the CMake curses interface, configure the build, generate makefiles, and build phish:
Note that if you add a new minnow to the minnow directory, simply re-run ccmake regenerate makefiles, and build - your minnow will automatically be incorporated into the build.
Alternatively, typing the following from the minnow directory will build all C and C++ minnows:
where machine is the suffix of one of the provided Makefiles, e.g. linux.mpi or linux.zmq. Type "make" to see a list of the different files and what compiler and MPI options they support.
The ".mpi" or ".zmq" suffix of the make target and associated Makefile refer to which version of the PHISH library will be linked against, either the MPI or ZMQ version.
The make command also builds non-PHISH C or C++ programs which are intended to be wrapped with one of the "wrap" minnows discussed above so they can be used as a minnow. Examples are the echo and reverse programs in the minnow directory.
If none of the provided Makefiles are a match to your machine, you can use one of them as a template and create your own. Note that only the top section for compiler/linker settings need be edited. This is where you should specify your compiler and linker and any switches they use. For the LIB setting, be sure to use the appropriate version of the PHISH library you are linking to, i.e. libphish-mpi.so or libphish-zmq.so.
IMPORTANT NOTE: When adding a new minnow that is a single file to the minnow directory, you should insure the string "MINNOW" appears somewhere in the *.cpp or *.c file. This is how the top-level minnow/Makefile includes it in the build list. It will then be automatically built with the other minnows.
IMPORTANT NOTE: If you wish to switch the PHISH library used with your minnows (MPI vs ZMQ), you should type "make clean-all" and then re-compile and re-link all the minnows. You can also type "make clean" to simply delete all object files.
If your new minnow is complex enough to consist of multiple files, you can add a specific rule for how to build it to the Makefile.machine you use, e.g. that defines a new target with a list of OBJ files that it depends on. Or you can build it in a separate directory with your own custom Makefile, so long as you link to the PHISH library, similar to how the Makefiles in the minnow directory perform the final link step.
Your executable minnow files do not need to be added to the minnow directory. See the --path command-line switch for the bait.py tool for how to access minnows from other directories when running a PHISH net.