Extending GlusterFS with Python

Are you a Python programmer who wishes your storage could do more for you? Here's an easy way to add functionality to a real distributed filesystem, in your favorite language.

Programming languages are usually not good neighbors. Even mixing languages as closely related as C and C++ often can lead to a morass of conflicting conventions with respect to symbol names, initialization orders and memory management strategies. As the distance between languages increases, the difficulty of integrating them increases as well. This is particularly true when attempting to mix compiled and interpreted languages. Most interpreted languages have ways to call functions and access symbols in compiled libraries, but these facilities often are far from convenient, and calling back the other way—from compiled code to interpreted—is less convenient still. Integration between interpreted languages is even less feasible—the one notable exception being the several languages that share the Java Virtual Machine (JVM). Interoperability between interpreted languages using different virtual machines usually is limited to message passing between separate processes.

In this context, Python's facilities for integrating with code written in other languages are like a breath of fresh air. One option is Jython, which exists quite comfortably within the aforementioned JVM ecosystem. For integration with compiled code, Python offers not one but two methods of integration. The first is the "extension API", which allows you to write Python modules in C. ("C" is used here as shorthand for any compiled code that adheres to the initialization and calling conventions originally defined for C.) Using this interface, it is possible to create compiled modules that offer the full functionality of native Python modules with the full performance of compiled code. There even are projects like Cython that will generate most of the necessary "boiler plate" for you.

The Python ctypes module offers an even more convenient option for integration with compiled code, with only a very small decrease in functionality. Using ctypes, Python code can call functions and access symbols even in C libraries whose authors never thought about Python at all. Python programmers also can use ctypes to interpret C data structures (overlapping somewhat with the functionality provided by the struct module) and even define Python callbacks that can be passed to C functions. Although it is not possible to do absolutely everything with ctypes that you can do with the extension interface, combining the two approaches can lead to very powerful results.

As a case study in combining Python code with an existing compiled program or language, this article focuses on the implementation of a Python "translator" interface for GlusterFS. GlusterFS is a modern distributed filesystem based on the principle of horizontal scaling—adding capacity or performance to a system by adding more servers based on commodity hardware instead of having to pay an ever-increasing premium to make existing servers more powerful. Development is sponsored by Red Hat, but it's completely open source, so anyone can contribute. In addition to horizontal scaling, another core principle of GlusterFS is modularity. Most of the functionality within GlusterFS actually is provided by translators—so called because they translate I/O calls (such as read or write) coming from the user into the same or other calls that are passed on toward storage. These calls are passed from one translator to another, arranged in an arbitrarily complex hierarchy, until eventually the lowest-level calls are executed on servers' local filesystems. I call this interface TXAPI here for the sake of brevity, even though that's not an official term. TXAPI has been used to implement internal GlusterFS functionality, such as replication and caching, and also external functionality, such as on-disk encryption.

This article is not primarily about GlusterFS, however. Even though I use GlusterFS to illustrate techniques for integrating Python and C code and show results to illustrate the potential benefits of such integration, most of the techniques are equally applicable to other programs with a similar set of characteristics. Those characteristics include a C "top level" calling into Python instead of the other way around, a fundamentally multithreaded execution model, and the presence of a well-defined plugin interface (TXAPI) that makes extensive use of callbacks in both directions.

The fact that GlusterFS is primarily a C program—filesystems are, after all, system software—means that you can't use ctypes for everything. To bootstrap your integration, you need to use Python's "embedding API", which is a close cousin of the previously mentioned extension API and allows C code to call in to the Python interpreter. You need to invoke this API at least once to create an interpreter and invoke an initialization function in a Python module. For this purpose, you use a single C-based "meta translator" that can be loaded just like translators always have been. This translator is called glupy from GLUster and PYthon. (The preferred pronunciation is "gloopy" even though "glup-pie" might make more sense given those origins.) Most of what glupy does is provide the generic embedding-API glue to load the actual Python translator, which is specified as an option. This loading is a fairly simple matter of calling PyImport_Import to load the module, followed by PyObject_CallObject to initialize it, as shown below (error handling has been left out for clarity):


priv->py_module = PyImport_Import(py_mod_name);
Py_DECREF(py_mod_name);

py_init_func = PyObject_GetAttrString(priv->py_module, "xlator");

py_args = PyTuple_New(1);
/* "this" is the C pointer to this glupy instance */
PyTuple_SetItem(py_args,0,PyLong_FromLong((long)this));

priv->py_xlator = PyObject_CallObject(py_init_func, py_args);
Py_DECREF(py_args);

The user's Python init function is then responsible for registering TXAPI callbacks for later, in addition to its own domain-specific initialization. Glupy also includes a Python/ctypes module that encapsulates the GlusterFS types and some functions that glupy users can invoke (in the example, this is done using the "dl" handle).

At this point, you reach a fork in the road. If you're already using the embedding API, why not continue using it for almost everything? In this approach, a glupy dispatch function would use Py_BuildValue to construct an argument list and then use PyObject_CallObject to call the appropriate Python function/method from a table. This is pretty tedious code to write by hand, but much of the process could be automated. The bigger problem with this approach is that TXAPI involves many pointers to GlusterFS-specific structures, which must be passed through the embedding API as opaque integers. The Python code receiving such a value must then explicitly use from_address to convert this into a real Python object. Clutter within glupy itself is not a problem, but clutter within glupy users' code makes this approach less appealing.

The approach actually used in glupy involves less C code and more Python code, with a greater emphasis on ctypes. In this approach, the user's Python code is presented not as Python functions but as C functions, using ctypes to define function types that then can be used as decorators. Unfortunately, details of the platform-specific foreign function interfaces used by ctypes to implement such a callback mean that there's no way to get the actual function pointer as it's seen by C code other than by actually passing it to a C function. Accordingly, you pass the Python callback object to a glupy registration function that can see the result of this conversion. For each type of operation, there are two corresponding registration functions: one for the dispatch function that initiates the operation and one for the callback that handles completion. The glupy meta-translator then stores pointers to the registered functions in a table for fast access later. One side effect of this approach is that glupy functions are strongly typed. This might seem rather un-Pythonic, but TXAPI itself is strongly typed, and the consequences of mixing types could be a hung filesystem, so this seems like a reasonable safety measure. Although this might all seem rather complicated, the net result is Python code that's relatively free of type-conversion clutter and requires very little initialization code. For instance, the following shows the init function for an example I'll be using that registers dispatch functions and callbacks for two types of operations:


def __init__ (self, xl):
        dl.set_lookup_fop(xl,lookup_fop)
        dl.set_lookup_cbk(xl,lookup_cbk)
        dl.set_create_fop(xl,create_fop)
        dl.set_create_cbk(xl,create_cbk)

The next problem to solve is multithreading. The Python interpreter still is essentially single-threaded, so C code that calls into Python must be sure to take the Global Interpreter Lock and do other things to keep the interpreter sane. Fortunately, current versions of Python make this much easier than it used to be. The first thing you need to do is enable multithreading by calling PyEval_InitThreads after Py_Initialize. What a surprising number of people seem to miss, even though it's fairly well documented, is that part of what PyEval_InitThreads does is acquire the Global Interpreter Lock on behalf of the calling thread. This lock must be released explicitly at the end of initialization, or else any other code that tries to acquire it will deadlock. In this case, this acquisition is implicit in calls to PyGILState_Ensure, which is the recommended way to set up interpreter state before calling into Python from multithreaded C code. Each glupy dispatch function and callback does this, with a matching call to PyGILState_Release after the Python function returns.

Before moving on from what's inside glupy to what glupy code looks like, you need to know what this example glupy-based translator actually does. The problem this example tries to solve is one that occurs frequently when using GlusterFS to store the code for PHP Web applications. Often, such applications try to load literally hundreds of include files every time a page is requested. Each include file might exist in any of several include directories along a search path. The example caches information about "positive lookups" (that is, those that succeeded) but not about "negative lookups" (which failed).

Although this behavior makes sense for many applications, the performance impact for many PHP applications can be severe. Without negative-lookup caching, you're likely to search half of those directories in vain before finding the one that contains each include file, every time the including page is requested. (This pattern does occur in other environments as well, including Python Web applications, but common PHP frameworks cause those applications to be hit the hardest.) Just as the effects are severe, the benefits of adding a negative-lookup cache can be significant. For example, a C version of such a translator decreased average include-search times nearly seven-fold. What could a Python version do?

Here's part of a translator based on glupy:


@lookup_fop_t
def lookup_fop (frame, this, loc, xdata):
      pargfid = uuid2str(loc.contents.pargfid)
      print "lookup FOP: %s:%s" % (pargfid, loc.contents.name)
      # Check the cache.
      if cache.has_key(pargfid) and (loc.contents.name in cache[pargfid]):
              dl.unwind_lookup(frame,0,this,-1,2,None,None,None,None)
              return 0
      key = dl.get_id(frame)
      requests[key] = (pargfid, loc.contents.name[:])
      dl.wind_lookup(frame,POINTER(xlator_t)(),loc,xdata)
      return 0

This is the function that gets called to look up a file, which is the core functionality for this example. Entry to this function represents a transition from C to Python, while its return represents a transition back to C. Calls through the "dl" object—a handle to the C dynamic library that supports glupy—also suspend the Python interpreter while they run. The Python decorator syntax allows you to hide most of the function-type details, and there's also a notable lack of type-conversion code. Most of what's there is domain-specific code, not boiler plate required by the infrastructure.

In the top half of this function, you simply check the cache to see if you already know the requested file won't be there. If the cache check succeeds, the lookup fails immediately, and you "unwind" the translator stack to report that fact. As with the registration functions, each operation type has its own specific wind (call downward) and unwind (return upward) functions as well. This represents a temporary return from the "Python world" to the "C world", and it's worth noting that these transitions between worlds might occur seamlessly many times while processing a single request. In particular, a common GlusterFS translator idiom is for a completion callback on one request to initiate the next, and if that request completes immediately (as done here), then you can have multiple requests and completions all on the stack at once.

Returning to the code, if you do not find an entry in the cache (and you already know it must not be in the standard positive-lookup cache or else you wouldn't even have been called), you pass the request on to the next translator using wind_lookup. When that next translator is done, it returns control (through the glupy meta-translator) to lookup_cbk. Here you retrieve your request context, conveniently stashed in a dictionary for you by lookup_fop, and use it to update the cache according to whether the file was found.

There are a few other less relevant details of how this particular glupy translator works, but that really is the meat of it. With less than a hundred lines of Python code, including comments and empty lines, you can add a significant piece of functionality to a real filesystem. But, how well does it really work? As it turns out, it works very well; see Table 1. A simple test reveals that the result is slower than the C-based version of the same thing, but still more than four times as fast as the baseline. Clearly, the fact that you're caching these results matters more than what language you're using to do it.

Table 1. Results of Caching Failed-Lookup Requests
ms/lookup minimum average maximum 99th percentile
no caching 0.368 6.898 16.286 9.702
C version 0.379 1.036 18.503 2.180
glupy version 0.381 1.527 21.163 2.916

As promising as these results are, they're more of a beginning than an end. Glupy is still a very young project, and much remains to be done. Support needs to be added for a few dozen more operation types and several data structures. There still are more ways that GlusterFS calls into translators and utility functions that translators themselves call. There are many ways the glupy interface could be made more convenient, and there are undoubtedly performance or concurrency issues still to be resolved. The most important thing is that the basic infrastructure for doing all of these things already exists, and not just for GlusterFS translators. If even a highly multithreaded and asynchronous program like this can take advantage of all that Python has to offer, so can just about any other program. Thanks to Python's extension/embedding interface and ctypes module, a "best of both worlds" approach to developing complex software is more achievable than most people think.

Resources

Jython: http://www.jython.org

Cython: http://www.cython.org

GlusterFS: http://www.gluster.org

Glupy Source Repository: https://github.com/jdarcy/glupy

Negative-Lookup Caching Translator in C: https://github.com/jdarcy/negative-lookup

Zend (PHP) Framework on Include Files: http://framework.zend.com/manual/1.12/en/performance.classloading.html

Load Disqus comments