Distributing computing with RPyC
A native Python option for seamless control of many machines
This content is part # of # in the series: Charming Python
This content is part of the series:Charming Python
Stay tuned for additional content in this series.
Back in 2002, I wrote a series of articles on
"distributing computing" (see Related topics). At that
time, RPyC did not exist, but I covered the Python tool Pyro and the
xmlrpclib. Surprisingly little has
changed in the last seven years. Pyro is still around and
actively maintained; and
xmlrpclib still exists
in Python, as does
these have been refactored as
xmlrpc.server in Python 3).
The space RPyC steps into is very much the one already occupied by Pyro and XML-RPC. RPyC does not really add anything fundamentally new to these longstanding tools, but there are some nice features to the design of RPyC.
RPyC 3.0+, in a nutshell, has two modes to it: a "classic mode," which was already available prior to its version 3, and a "services mode," which was introduced in version 3. The classic mode lacks any real security framework (which isn't always a bad thing) and simply presents remote machines as if they were local resources. The newer, services, mode isolates a collection of published interfaces that a server supports and is secure inasmuch as it prohibits everything not explicitly permitted. The classic mode is essentially identical to Pyro (without Pyro's optional security framework); the service mode is essentially RPC (for example, XML_RPC), modulo some details on calling conventions and implementation.
Background on distributing computing
In the paradigm of stand-alone personal computing, a user's workstation contains a number of resources that are used to run an application: disk storage for programs and data; a CPU; volatile memory; a video display monitor; a keyboard and pointing device; perhaps peripheral I/O devices such as printers, scanners, sound systems, modems, game inputs, and so on. Personal computers also have network capabilities, but conventionally, a network card has largely been just another sort of I/O device.
"Distributing computing" is a buzz-phrase that has something to do with providing more diverse relationships between computing resources and actual computers. In the old days, we used to speak of "client/server" and "N-Tier architecture" to describe hierarchical relationships among computers. However, different resources can enter into many different sorts of relationships—some hierarchical, others arranged in lattices, rings, and various other topologies. The emphasis has shifted towards graphs, away from trees. Some of the many possible examples:
- SANs (storage-area networks) centralize persistent disk resources for a large number of computers.
- In the opposite direction, peer-to-peer (P2P) protocols such as Gnutella and Freenet decentralize data storage and its retrieval.
- The X Window System and VNC (AT&T's Virtual Network Computing) allow display and input devices to connect to physically remote machines.
- Protocols such as Linux Beowulf allow many CPUs to share the processing of a complex computation, whereas projects such as SETI@Home (NASA's Search for Extraterrestrial Intelligence), GIMPS (Great Internet Mersenne Prime Search) and various cryptographic "challenges" do the same with much less need for coordination.
- Ajax is another means, albeit specifically within a Web browser client, of utilizing resources from many sources.
The protocols and programs that distribute what were basically hardware resources of old-fashioned PC applications make up only part of the distributed computing picture. At a more abstract level, much more interesting things can be distributed: data, information, program logic, "objects," and, ultimately, responsibilities. DBMSs are a traditional means of centralizing data and structuring its retrieval. In the other direction NNTP, and later P2P, radically decentralized information storage. Other technologies, such as search engines, restructure and recentralize information collections. Program logic describes the actual rules of proscribed computation (various types of RMI and RPC distribute this); object brokerage protocols such as DCOM, CORBA, and SOAP recast the notion of logic into an OOP framework. Of course, even old-style DBMSs with triggers, constraints, and normalizations always carried a certain degree of program logic with them. All of these abstract resources are at some point stored to disks and tapes, represented in memory, and sent as bitstreams over networks.
In the end, what is shared between distributed computers are sets of responsibilities. One computer "promises" another that under certain circumstances it will send some bits that meet certain specifications over a channel. These promises or "contracts" are rarely first about particular configurations of hardware, but are almost always about satisfying functional requirements of the recipients.
Why RPyC is useful
Before outlining what RPyC does, let me suggest three categories of resources/responsibilities that RPyC can be used to distribute:
- Computational (hardware) resources. Some computers have faster CPUs than others; and some likewise have more free cycles on those CPUs once the process priorities and application loads are considered. Similarly, some computers have more memory than others, or more disk space (important, for example, in certain large-scale scientific calculations). In some cases, specialized peripherals might be attached to one machine rather than another.
- Informational resources. Some computers may have privileged access to certain data. Such privilege can be of several sorts. On the one hand, a particular machine might be the actual originating source of data; for example, because it is attached to some sort of automated data collector such as a scientific instrument or because it is a terminal into which users enter data (a cash register, a check-in desk, an observation site, etc.). On the other hand, a database might be local to a privileged computer, or at least to a limited group that the machine or account belongs to. Non-privileged computers might nonetheless have reason to have access to certain aggregate or filtered data derived from the database.
- Business logic expertise. Within any organization—or between organizations—certain parties (individuals, departments, etc.) have the capacity and responsibility to decide the decision rules in certain domains. For example, the payroll department might determine (and sometimes modify) the business logic concerning sick days and bonuses. Or Jane, the database administrator, might have the responsibility to determine the most efficient way to extract this datum from complex relational tables.
RPyC lets you distribute all those resources.
RPyC classic mode
The RPyC classic mode, in essence, just lets all the capabilities of a
remote Python installation run within a local Python system. Security here
basically amounts to the same thing as giving a shell account to whoever
connects. If security is at issue, the encryption library
tlslite can encrypt connections and require
logins for connections. In other words, you can easily create the
equivalent of ssh rather than the equivalent of telnet. Of course, the
advantage here is that these connections can be controlled by Python
scripts, and allow more robust interaction between local and remote
resources than does a language such as Expect.
To launch a server on a remote machine, simply run
classic_server.py that comes with RPyC. If you
want a secured connection, add the
option. To customize a port, use
--help for additional options in the
server. You can, of course, launch your server as part of general system
initialization or in cron jobs, to make sure they are running on a given
machine. Once some servers are running, you can connect to them from as
many clients as you like. Note, however, that servers are not really
that special; the same machine and process can utilize many
servers and act as server to many clients at the same time (including two
machines symmetrically "serving" each other).
A shell session (once some servers are launched) shows this:
Listing 1. System and Python information
>>> import sys,os >>> os.uname() # Some info about the local machine ('Darwin', 'Mary-Anns-Laptop.local', '10.0.0d1', 'Darwin Kernel Version 10.0.0d1: Tue Jun 3 23:40:01 PDT 2008; root:xnu-1292.3~1/RELEASE_I386', 'i386') >>> sys.version_info # Some info about the local Python version (2, 6, 1, 'final', 0) >>> os.getcwd() '/Users/davidmertz'
Now, let's import RPyC and connect to a few servers. Before doing this, I had launched local servers from two different Python versions, and one on a remote machine.
Listing 2. Importing RPyC and connecting to servers
>>> import rpyc >>> conn26 = rpyc.classic.connect('localhost') >>> conn26.modules.os.uname() ('Darwin', 'Mary-Anns-Laptop.local', '10.0.0d1', 'Darwin Kernel Version 10.0.0d1: Tue Jun 3 23:40:01 PDT 2008; root:xnu-1292.3~1/RELEASE_I386', 'i386') >>> conn26.modules.sys.version_info (2, 6, 1, 'final', 0) >>> conn26.modules.os.getcwd() '/Users/davidmertz/Downloads/rpyc-3.0.3/rpyc/servers' >>> conn25 = rpyc.classic.connect('localhost',port=18813) >>> conn25.modules.os.uname() ('Darwin', 'Mary-Anns-Laptop.local', '10.0.0d1', 'Darwin Kernel Version 10.0.0d1: Tue Jun 3 23:40:01 PDT 2008; root:xnu-1292.3~1/RELEASE_I386', 'i386') >>> conn25.modules.sys.version_info (2, 5, 1, 'final', 0) >>> conn25.modules.os.getcwd() '/Users/davidmertz/Downloads/rpyc-3.0.3/rpyc/servers' >>> connGlarp = rpyc.classic.connect("22.214.171.124") >>> connGlarp.modules.os.uname() ('FreeBSD', 'antediluvian.glarp.com', '6.1-RELEASE', 'FreeBSD 6.1-RELEASE #0: Fri Jul 18 00:01:34 MDT 2008; email@example.com:/usr/src/sys/i386/compile/ANTEDILUVIAN', 'i386') >>> connGlarp.modules.sys.version_info (2, 5, 2, 'final', 0) >>> connGlarp.modules.os.getcwd() '/home/dmertz/tmp/rpyc-3.0.3/rpyc/servers'
You can see that we have connections to a couple of different machines with
different Python versions. The functions and attributes we accessed are
arbitrary, but the important point is that we might call any
functions or classes available on those machines. So, for example, if I
know the machine
antediluvian.glarp.com has a
Payroll installed on it that has
get_salary() in it, I might
Listing 3. Calling get_salary()
Over at antediluvian, there might be a local database installed, or it might even make its own connections to other resources. What is returned by my function call, however, is simply the same data that would be returned if the function was run locally on antediluvian.
Putting code on the remote machine
Running standard module functions on a remote machine is a cute trick, but what we often want to do more usefully is run our own code remotely. There are several ways of doing this within RPyC's classic mode. The most direct way is perhaps to simply open a Python shell onto that machine via the connection we have established. For example:
Listing 4. Opening a Python shell on a remote machine
>>> conn = rpyc.classic.connect('linux-server.example.com') >>> rpyc.classic.interact(conn) Python 2.5.2 (r252:60911, Oct 5 2008, 19:24:49) [GCC 4.3.2] on linux2 Type "help", "copyright", "credits" or "license" for more information. (InteractiveConsole) >>> #... run any commands on remote machine
From such a remote shell, you can define whatever functions or classes you want, import any outside code, or do anything you can do from a local Python shell.
There is also an RPyC function called
that seems to send a local object over to a remote server, where
presumably it might run in the local context. Unfortunately, I was not
able to get this function to act as I expected (I am sure I just have some
syntax wrong, but the documentation is vague). As a kludge, you can
directly execute (or
eval()) code in the remote
server. For example:
Listing 5. Executing code on a remote machine
>>> # Define a function (or class etc) as actual source code >>> hello_txt = """ ... def hello(): ... import os ... print "Hello from", os.getcwd() ... """ >>> exec hello_txt # Run the function definition locally >>> hello() Hello from /tmp/clientdir >>> conn.execute(hello_txt) # Run the function definition remotely >>> remote_hello = conn.namespace['hello'] >>> remote_hello() # Displays on remote server >>> with rpyc.classic.redirected_stdio(conn): # Redirect to local client ... conn.namespace['hello']() ... Hello from /tmp/serverdir
What we have done is compile a function into the remote namespace and run
it as if it were a local function. However, we need to also grab that
remote console output if we want to actually see the output from its
had instead returned a value, though, it would still be returned to
the local context as in the prior examples.
What we did with the
with context above was
basically a kind of "monkey patching." That is, we temporarily used a
local STDIO in place of a remote STDIO. RPyC lets you do this in general
with system resources. Code might run (that is, use CPU and memory) on a
local machine but still use some key resource on the remote machine. That
might be a computationally expensive function call if the server has more
CPU resources, but often it is something such as a socket where the remote
machine has different access domains. For example:
Listing 6. Monkey patching a socket connection
import myserver c = rpyc.classic.connect('machine-outside-firewall') myserver.socket = c.modules.socket # ...run myserver, which will now open sockets outside firewall
If utilizing a resource on a remote machine is time consuming, the local program using that resource might seem to stall while the remote action is completing. However, this need not be so if you make these calls asynchronous. A remote object can report whether it has a result ready, and you can do other local actions in the meantime (or perhaps actions utilizing other remote servers).
Listing 7. An asynchronous call
>>> conn.modules.time.sleep(15) # This will take a while >>> # Let the server do the waiting >>> asleep = rpyc.async(conn.modules.time.sleep) >>> asleep async(<built-in function sleep>) >>> resource = asleep(15) >>> resource.ready False >>> # Do some other stuff for a while >>> resource.ready True >>> print resource.value None >>> resource <AsyncResult object (ready) at 0x0102e960>
In our example
resource.value is uninteresting.
However, if the remote method that we made asynchronous returned a value,
that value would be available once
resource.ready became True.
RPyC service mode
I will write relatively little about the newer RPyC service mode. The classic mode is really the more general system, even though it is itself built as a new-style service in just a few lines of code. A service under RPyC is really little different from XML-RPC (or whatever-RPC). Classic mode is just a service that exposes everything on the remote system, but you can build services that expose just a few things in a small number of lines of code. For example:
Listing 8. Using service mode to expose a few methods
import rpyc class DoStuffService(rpyc.Service): def on_connect(self): "Do some things when a connection is made" def on_disconnect(self): "Do some things AFTER a connection is dropped" def exposed_func1(self, *args, **kws): "Do something useful and maybe return a value" def exposed_func2(self, *args, **kws): "Like func1, but do something different" if __name__ == '__main__': rpyc.utils.server.ThreadedServer(DoStuffService).start()
From a client, this service is just like the classic mode server, except
all it exposes is the methods that are prefixed by
exposed_ (minus the prefix). Trying to access
other methods (such as builtin modules) will fail. So a client might look
Listing 9. Calling service-exposed methods
>>> import rpyc >>> conn = rpyc.connect('dostuff.example.com') >>> myval = conn.root.func1() # Special 'root' of connection >>> local_computation(myval)
There is more in RPyC than I have mentioned. For example, like Pyro, RPyC provides a "Registry" that lets you name services and access them by name rather than by domain name or IP address. This is straightforward, and the RPyC documentation explains it.
As I have indicated in this article—and as RPyC's own documentation says
explicitly—RPyC is in many ways "Yet another RPC package." For example,
the service we briefly constructed above is nearly identical to the same
code we would write using
The only difference is the wire protocol used for requests and results.
Nonetheless, despite some minor glitches I encountered, RPyC is well
constructed and very simple to get running. You can easily construct a
robust distribution of resources and responsibilities using just a few
lines of RPyC code.
- Read David's series of "Distributing Computing" articles:
- The RPyC home page contains good documentation, as well as the downloads and links to screencasts and discussion areas.
- In the developerWorks Linux zone, find more resources for Linux developers, and scan our most popular articles and tutorials.
- See all Linux tips and Linux tutorials on developerWorks.