PHISH Pheatures

• The PHISH package is open-source software, distributed under the Berkeley Software Development (BSD) license. This effectively means that anyone can use the software for any purpose, including commercial redistribution.
• The PHISH library is a small piece of code (few 1000 lines), with a compact API (couple dozen functions). It has a C interface, so that it can be easily called from programs written in a variety of languages (C, C++, Fortran, Python, etc). The library can be compiled on any platform with a C++ compiler.
• The PHISH library comes in two flavors with the same API: one based on message passing via the MPI library, the other based on sockets via the open-source ZMQ library. This means you need one or both of these packages (MPI, ZMQ) installed on your machine to build a minnow (process) that uses the PHISH library.
• PHISH nets (programs) consist of one or more collections of minnows (schools), hooked together in defined communication patterns, to encode an algorithm or computation, The toplogy of a PHISH net is specified in an input script, which is text files with a simple command syntax.
• PHISH minnows can define one or more input and output ports for sending and receiving datums. This allows schools of minnows to be hooked together in a variety of communication patterns.
• PHISH minnows can be written to operate on datums of various types (e.g. integers or floating-point values or strings). This allows minnows to be re-used in various PHISH nets.
• A PHISH wrapper (pun intended) for Python is provided, so that minnows that call the PHISH library can be written in Python. Minnows written in different languages (e.g. C++ or Python) can be used interchangeably in a PHISH input script.
• The PHISH library exchanges data between minnows with strict data typing rules, so that minnows can be written in different languages (e.g. C++ vs Fortran vs Python) or run on different machines (4-byte vs 8-byte integers).
• PHISH input scripts are processed via a tool called bait.py tool which can also run them as an MPI or socket-based program.
• PHISH input scripts use a hook command which allows data to be exchanged in various patterns between schools of minnows. This enables parallelism in data processing to be easily expressed and exploited.
• PHISH nets can be run on a single processor, so long as the OS supports multiple processes. They can be run on a multicore box. They can be run on any distributed-memory or shared-memory platform that supports MPI or sockets. Or they can be run on a geographically dispersed set of machines that support socket connections.
• A PHISH net can look for incoming data on a socket port. It can likewise export data to a socket port. This means that two or more PHISH nets can be launched independently and exchange data. This is a mechanism for adding/deleting minnows (processes) to/from a calculation on the fly.
• PHISH minnows are included that wrap non-PHISH applications that read from stdin and/or write to stdout. This allows such an application to be used in a PHISH net and exchange data with other minnows.