This document provides information for the W3C members and other interested community. This document does not specify a W3C standard of any kind.
Feedback should be directed to the author.
A list of current W3C documents can be found at: http://www.w3.org/pub/WWW/TR
The Common Gateway Interface (CGI) is not scaling to meet the requirements of today's dynamic, interactive webs. For this reason, multiple vendors have proposed C callable APIs. These APIs allow authors to alleviate the performance penalty of CGI, and allow tighter integration of add-in modules. Unfortunately, this comes at the price of complexity and portability.
This document describes a new model for extending WWW servers. First, HTTP is captured using an interface specification, which eliminates the ambiguities of interpretating a standards-track document. This interface is then implemented atop a particular httpd's API. Finally, all of this is done using a standard distributed object model called ILU.
Digital Creations' work on our ILU Requester reflects this design and shows its advantages. This paper describes the ILU Requester.
Applications deployed over the World-Wide Web often involve an HTTP server integrated with a legacy information system, or a custom information system. The Common Gateway Interface, or CGI, is the most widely deployed mechanism for integrating HTTP servers with other information systems, but studies have shown that its design does not scale to the performance demands of contemporary applications. Microsoft states that applications for their API are five times faster than CGI applications.
Moreover, CGI applications do not run in the httpd process. In addition to the performance penalty, this means that CGI applications cannot modifiy the behavior of the httpd's internal operations, such as logging and authorization. Finally, CGI is viewed as a security issue by some server operators, due to its connection to a user-level shell.
A current solution is to use an httpd with an API, such as Apache 1.x or Netscape. By using the API, you have a performance increase and a load decrease by running your application in the httpd process, rather than starting a new process for every request. Also, the API exposes some of the httpd's own behavior, allowing you to modify its operation. In fact, servers like Apache implement large portions of their functionality, such as ISMAP handling and logging, as modules.
Unfortunately, the API benefits come at a price. Running a user-written module inside the httpd process leads to possible reliability concerns. For instance, when developing our requesters, early code would regularly lead to core dumps from unhandled errors, as well as memory leaks. Also, most current servers use either multiple pre-forked subprocesses or separate threads for each new request. Thus, applications which change state, such as a simple counter script, have data concurrency issues that are the burden of the programmer to solve.
Most importantly, the API route eliminates the casual CGI programmer. In a recent survey, Perl beat C 4 to 1 with 46% of the total votes. It appears that the possibilities for language-choice in a C-based API mechanism are restrictive.
Finally, the portability of CGI applications from one httpd implementation to another, would be lost with an API strategy. Since each API has a different syntax, authors would be forced to know each API beforehand. Thus, APIs could become instruments used by vendors to ensure market retention.
The elimination of scripting by an API strategy is a serious issue. Web services are usually built using scripting languages such as Perl, Python, Tcl, Visual Basic, Rexx, etc. This seems to be the case because web apps are frequently:
In essence, the genre of CGI applications are usually complex enough to use tools good for rapid prototyping, but which rarely get past the prototype stage and into C.
To address this next generation of server-extending, we developed a mechanism based on a uniform interface specification for HTTP. This is the HTTP.isl. By basing our extension mechanism on a distributed object protocol like ILU, we get the performance and features of an API strategy (as shown below), with the portability and simplicity of CGI. Moreover, it permits the httpd to be extended not only out of its address space, but off its machine, and thus into capabilities available only on a remote node. This is in the true client-server fashion.
We call this extension mechanism the ILU Requester.
We have listed the problems with the current CGI/API situation. And we have described an ILU Requester architecture. What are its requirements, and what are some preferred possibilities?
We have implemented the ILU Requester for several platforms, and have extended development to include other interested parties. First, we will give some background, and then a description.
In December of 1994, we were tasked with developing a complex WWW service. This service necessitated a dynamic language, and had state. Yet, we were forced to use CGI. Thus, we made a first implementation using a long-running process that managed the state using a dynamic language (Python), and a small "controller" script that would message it on each hit.
Over time, we found that we were inventing our own client/server protocol. For this and other reasons, we started looking at using ILU to manage interactions between processes. Thus, the CGI script got a surrogate reference to an encapsulation of the stateful system.
Still, we had the performance penalty of CGI. In April of 1995, we wrote a patch for Apache 0.6.5 that embedded the ILU runtime. With this, we had access to objects via registered URL constructs. This served several production systems into the fall. At this point, we started to refer to this embedded ILU module as the requester.
In August, a version of Apache was released that had an API, so we started reworking the requester to use it. By October we had a related requester for Netsite working on Unix and partially on NT. In December, based on a new draft of the HTTP spec, we consolidated the two feature sets, and wrote an HTTP ISL that was comprehensive with respect to the new specification. Also, we started work with ILU 2.0.
In January of this year, we started standardizing on a Python "framework" module for creating our online services. For this, we developed an ISL for installing object-based Authorizers and Loggers "into" the httpd.
Also in January, we released our first major product based on this architecture called Broadcast. This is a Web-based chat application that had one primary goal: it should be very fast under very highly-loaded conditions. Some design choices were:
We chose to use an ILU Requester that would make generic calls on published objects that represented the chat site's components. This allowed us to have very low latency (by avoiding startup costs), and expose the OO design of the chat implementation.
It appears that the design choice was valid. Performance is fantastic, and load is low. Also, using the requester strategy, we now have many new possibilites for application partitioning. Finally, using our API Scripting infrastructure, we are able to add new features in a very coherent fashion.
As alluded to above, the entire system is based around ILU. From this, we get language-independence, cross-process communication, and platform-independence.
The goal is to add an abstract object interface to an httpd in a uniform way. For this, we wrote an interface for HTTP that encapsulates the behavior of the HTTP transaction. We then implement this interface in C by mapping it to the semantics of a particular httpd's API. This implementation is called the requester, and gives an httpd a mechanism for passing certain incoming requests to an ILU published object.
This architecture mimics the interaction between the browser and the httpd using the same concepts as HTTP. For instance, the information contained in the request is mapped into a Request type in the interface specification. The requested object is a Resource, and the result of the operation on the Resource is a Response. Both of the Request and Response are types defined in the interface.
Fortunately, because of the uniform, abstract ISL, the services you write do not have to know anything about the semantics of the server or its API. In fact, it would be possible to skip the httpd altogether, and communicate directly with the published object. The objects could do this: they can have multiple representations, and can communicate via HTTP requests, ILU requests, or some other request structure.
When writing a service, therefore, all you have to do is publish an object that is based on the skeleton code generated from the interface. Pretty standard stuff here. Then, if the published object is listed in the httpd's configuration file, incoming requests matching a certain URI form will be sent to the requester, which will make an ILU call to the published object.
Also, it is possible to map the requester to remotely-published objects using ILU's String Binding Handle mechanism. This makes it possible to bridge the httpd into services available on other platforms. Future ILU mechanisms will make this process easier.
As of this writing, we have solid requesters based on ILU 1.8 and 2.0 for Netsite (Unix) and Apache. They have been tested by others, reviewed for optimizations, passed through simple memory leak testers, and documented. We are making distributions freely available in source, and some in binary form. Currently, the requesters are known to work fine on Solaris, Digital Unix, Linux, AIX, and BSDI. Additionally, we have preliminary support for NT. Full support is waiting for us to finish up our work on threading with ILU. Finally, ILU has been reported to work on OS/2, and there is work on and Apache implementation for that platform.
The threading issue will become increasingly important as we build more sophisticated systems, especially when we might want to have a common ORB. However, for systems such as Netsite and Microsoft's IIS on NT, as well as Spyglass' server, it is required in the requester. This is because these platforms service each incoming request as a thread, rather than passing the request to an isolated process. Thread-safeing the requester is thus becoming a requirement.
We have just added support for aliasing multiply-published objects inside the requester. For instance, you could make a request to info@system, and have "system map" to one of several published objects. This is mainly for a performance increase in read-only situations. Note that this may be subsumed by the ILU work on multicast.
Another area we are working on is making the object systems easier to use. We have just added an HTTP header that gets returned, stating the ILU version and requester version. We are adding support for a discoverable interface, using a standard 'info@root' that is built in to every requester. This object will return catalogue information from config file directives, and will attempt to contact the ILU servers listed in the config file and get their 'info@root' information, if implemented.
Making an httpd able to call distributed objects is only half of the system: you must have objects that can be called.
We have intended for this system to replace CGI as a server-extension mechanism. To do this, it must be nearly as easy to create services as CGI is currently. For this, we have been working on an infrastructure for publishing requester-capable objects called API Scripting.
For creating services, we are focusing on Python, and building up a toolset to of components. We have made parts of this toolset available, and have released our demonstration programs and load testing modules. Based on this Python toolset and the requester, we are fielding high-performance Internet services for commercial use.
For instance, here is a very simple script in Python that publishes an object which echoes the contents of a request:
#!/usr/local/bin/python """Every good module deserves docstrings. This is a very simple script that subclasses a Resource, fills in the blanks, and echoes incoming Requests. It then publishes the object and goes into a main loop. """ import ilu import wwworb # our toolset class ILUforDummies(wwworb.Resource): def GET(self,request,connect): request = wwworb.Request(request) response = wwworb.Response(`request`) return response POST = GET # Create an ILU server ilu.CreateServer('paul.demos') # Now, create an instance of your class, passing it a parameter # for the name of the published object. nitwit = ILUforDummies('dumb') print nitwit.IluSBH() ilu.RunMainLoop()
Note that there are really only nine necessary lines in the above. This should put it into the realm of CGI for ease of use.
Our next step is to make highly-concurrent systems available in Python. To do this, we are working with the ILU team to thread the iluPrmodule. This work is related to the work on threading the ILU kernel.
For all of these, we have an emerging development group, and an infrastructure for documentation, tutorials, bug reports, etc.
Currently, we have stabilized our HTTP interface, and feel that it accurately represents the interaction between a browser and a client in a way useful for published objects behind an httpd. Therefore, we are now focusing on problem-specific interfaces.
First, we would like to have a discoverable interface for online services. For instance, one should be able to go to any requester-enabled site, send a request to an info@root published object, and get an inventory of that site. The contents of this inventory might vary, might support a set of minimum operations, might extend, and might change. All the things that an interface allows you to do over time.
This discoverable interface is being worked on. There are other interfaces that have already rolled out.
Most of our "API scripting" services involve persistent Python objects that receive requests from the ILU main loop. Some of these services need some type of Access Control List (ACL) mechanism on them. However, we really don't want to interface into some external, httpd-controlled, single-filesystem-based password file.
The API-based servers have modules already that allow you store user authentication information in a SQL table. Yet, we already have users defined in our object system. Moreover, we might want to have some instance-based authorization mechanism.
To extend the ACL-capabilities of the httpd, we wrote an Authorizer interface, and implemented it into the APIs we support. Thus, accesses to protected URIs are mapped to an object call, which determines if that operation is allowed by that identity.
Another area we wanted to standardize on was an intelligent logging mechanism. Currently, there is the Common Log File format for writing to disk. However, we wanted something more structured and more dynamic. Thus, we wrote an interface for logging, mapped it into the httpd's API functions, and created an installable logging facility. A smart implementation could publish an object which is registered for successful or unsuccessful requests to document or ILU-based requests. If there is an error, you could decide whether send a page to someone's beeper. For all requests, you can take the incoming data structure, and write parts of it into a miniSQL table.
We wanted to extend the httpd's logging facilities in new and interesting ways. For instance, we wanted to do processing and take special actions if an error was raised. Also, we wanted to investigate logging from a Unix httpd into a Windows-based personal DBMS like Microsoft Access.
To do this, we made an interface for loggging, and implemented it on the APIs we support. This mapping forces the httpd to run our object call during logging events. The interface is very simple; it just sends the Request object to LoggerObject via an asynchronous method. One could then subtype from there to do more interesting, platform-specific things.
The Stanford Digital Library team has produced interfaces and implementations for CORBA-type Common Object Services (COS). Common Object Services are objects or groups of objects that provide the basic requirements which most objects need in order to function in a distributed environment. These services are designed to be generic; they do not depend on the type of client object or type of data passed. Note: this is hard to do in ILU since there is no concept of the Object or Any type.
There are other good candidates for interfaces. For instance, the Harvest system has its own protocol for collecting indexing information, and doing searches. If an interface was written, it could perhaps be moved into this architecture.
We have started on some other standard interfaces, such as a Data Access interface and a an OLE interface (via Python). These, though, are not necessarily related to the ILU Requester, and are thus outside the scope of this paper.
It should be apparent that the architecture lends itself to good performance. However, we felt that some performance numbers were important, so we came up with a performance-testing program, and a regimen to exercise it.
To test, I used a Sparc 5 running Solaris 2.4 with 32 Mb as the testing client, and an Alpha with 64 Mb running Digital Unix as the testing server. The test program was written in Python, and used the httplib and thread modules to make concurrent requests. The server was running Netsite, using our requester and ILU 2.0. We had Netsite configured to use up to 32 processes.
We then ran through a series or URLs (listed below) in a series of two tests: a latency test and a throughput test. The latency test sent a series of requests on one thread, to test the response time. The throughput test dispatched the same number of requests on several simultaneous threads, to test concurrent use. Thus, the throughput test attempted to detail the affects of load, concurrency, and aggregate response time for a batch of requests.
For the URIs, the index.html test merely retrieved a very short HTML file. The others were:
simple.sh --------- #!/bin/sh echo 'Content-type: text/html\n\n' echo Hello. simple.pl --------- #!/usr/local/bin/perl print "Content-type: text/html\n\n"; print "hello."; simple.py --------- #!/usr/users/paul/cgipython """Simple script to echo the dictionary back. """ print 'Content-type: text/html\n\n' print 'Hello.' simple1.py ---------- #!/usr/users/paul/cgipython """\ Simple script to echo the dictionary back. """ print 'Content-type: text/html\n\n' import simple_lib simple_lib.py ------------- #!/usr/users/paul/cgipython """\ Simple script to echo the dictionary back. """ import cgi f = cgi.SvFormContentDict() print f.items() dumb.py ------- # Note: echo is equivalent to simple.py, and dumb is equivalent to simple1.py """\ The simplest, dumbest API script around. This Python program has one goal: fewest lines for an interactive script. The script reads the form variables, and sends them back, without very much formatting. Note that we have embedded the HTML into the class, which has added some characters. Normally this class would be even shorter, as we would use the "pyhtml" external representation. But, that would be smart, and this one is, well, dumb. """ import sys sys.path.append('wwworb') sys.path.append('interfaces') import string, ilu, wwworb print ilu.Version # Make a class derived from the Resource class in wwworb. Remember # that the base class (wwworb.Resource) requires a parameter to be # passed to its __init__ startup call. This parameter is the name # of the published object. class EchoforDummies(wwworb.Resource): def GET(self,request,connect): return wwworb.Response('Hello.') POST = GET class ILUforDummies(wwworb.Resource): def GET(self,request,connect): request = wwworb.Request(request) response = wwworb.Response(`request`) return response POST = GET # Create an ILU server ilu.CreateServer('paul.demos') # Now, create an instance of your class, passing it a parameter # for the name of the published object. nitwit = ILUforDummies('dumb') echo = EchoforDummies('echo') print nitwit.IluSBH() print echo.IluSBH() ilu.RunMainLoop()
These scripts were chosen to reflect both the least that could be done with a CGI script (echo back a string) vs. very little that could be done (parse the incoming request into data structures, and echo it back). The Bourne shell script and the Perl script are thrown in as reference points. It is the comparison of Python scripts that is relevant.
The Python interpreter used for the CGI scripts was very small. I removed nearly everything from the Modules setup, and did not link with threads (a source of startup time problems on Digital Unix). I used Python 1.3 for all of these, and ILU 2.0a3.
Thus, the comparison is between Python CGI and Python "API Scripting". The two tests are a simple echo of a string, and a slightly-computational parsing of the incoming information. Obviously, a real-world application, where files have to read, or marshals loaded, or databases connected-to, would tilt the scales towards API scripting, since the state is always in memory.
In the following, HPS refers to hits per second, SPH refers to seconds per hit, and SD refers to standard deviation.
This test used 10 runs of 1 thread, 20 requests on the thread:
This test used 10 runs of 10 threads, 20 requests apiece. In this case, the Min and Max refer to the thread completion times:
Understand that the HPS and SPH numbers on the throughput test reflect the ability of the server to service multiple requests simultaneously. Thus, each hit effectively is done faster.
Looking at the latency tests, you see that HTML and the simple CGI scripts are about the same HPS. These simple scripts don't parse the environment, and thus do no calculation. The simple1.py script which does parse the environment and imports a module suffers a 50% rise in latency. Yet, the API Scripting apps stay at the same level as the HTML and simple CGI, even though one is parsing the environment.
In the throughput test, the reference point -- the index.html file -- shows that a 10-thread request gets just over a five-fold bump in throughput. Certainly not a ten-fold, but a enough to show that it is handling simultaneous requests well. However, the CGI scripts start to show less benefit. Yet, the API Scripting applications stay at the
If we consider getting an HTML file -- both in single-threaded and ten-threaded batches -- to be a baseline, we see the relation of these tests. Again, we see that getting an HTML file gets a four-fold bump from a ten-thread batch. A simple Bash CGI script yields a three-fold improvement (317 percent) over single-thread HTML batches. A simple CGI script that parses the environment, run in ten-threaded batches, achieves only half the aggregate throughput of a single-threaded HTML request. Thus, concurrent CGI is slower than single-request HTML. Again, the API Scripting applications keep pace with the baseline.
1-thread % of
10-thread % of
10-thread % of
In the rightmost column above, which is a throughput measurement, an API Scripting application is over eight times faster than an equivalent CGI application.
A conclusion is that, even for simple state applications of reading in the form data, CGI loses to API Scripting in latency, and loses significantly in concurrent use. It would appear that the performance win would increase even more for complex applications, especially those that have to initialize some state, or make a connection to a SQL database. Getting the state setup for these is more complicated, and the increase in latency and load mean pileups for service.
A caveat in the testing must be noted. A more representative sample of API scripting vs. HTML would be to use an ILU C program and an API C program. This would also allow the testing of HTML vs. CGI vs. straight API C apps vs. ILU Requester with objects written in C.
At this time, movement to ILU 2.0 is the biggest issue. First, there are some minor bugs with the current prerelease. The real issue is embracing some new capabilities:
For more on this, see
We have a number of directions we intend to pursue internally, and a suggested direction for industry adoption.
Some of our plans are:
Some optional support:
Many ideas have floated around. Press releases have discussed, for instance, embedding Java inside of Web servers as a better fit than APIs. While this does get many of the benefits of this architecture, it is language-based, and thus does not have language-independent interfaces. Some, though, view this as a benefit.
I wanted to have a good language for parallel and distributed computing. For those purposes, however, the classic Java has very poor functionality. I like Java because it's simple and easy. But the basic idea of Java is not far from C++. C++ can also make objects, threads, and sockets. Java has no direct support for distributed object processing as C++ does not. So I decided to make a new framework for parallel and distributed computing.
Also in the FAQ, a comparison of HORB to CORBA:
CORBA and CORBA2 are desinged for Interoperability between different languages and different systems. You have to write interface definitions in CORBA IDL language in addition to real code. It must be annoying for casual use. CORBA cannot pass instances. It limits programming. CORBA ORB tends to huge to comply the CORBA standard. Since HORB ORB for clients is only 20KBytes, modem users can wait for dynamic loading. Current CORBA systems are very expensive. HORB is free of charge.
As stated on a demo page, HORB aims to "replace CGI or socket programming with smart remote object operations of HORB".
The interface for HTTP is used to extend the WWW server by mapping the browser-server interaction to an object request. We used the latest HTTP specification, as mentioned in the comment.
(* $Id: WD-ilu-requestor-960307.html,v 1.4 1996/03/07 17:02:48 connolly Exp $ *) (* Proposed HTTP interface Digital Creations <firstname.lastname@example.org> Reference: http://www.w3.org/pub/WWW/Protocols/HTTP1.0/draft-ietf-http-spec.html *) (* The following is a list of headers guaranteed to be included with the request, regardless of the requester used. This list is probably incomplete and will grow as I become more familiar with requesters other than NetSite: In "request.headers": None In "connection": "remote-ip" == the IP address of the remote client "remote-name" == the name of the remote clinet, or the IP address if the name cannot be determined *) INTERFACE http; TYPE field-name = ilu.CString; TYPE field-value = ilu.CString; TYPE optional-field-value = OPTIONAL field-value; TYPE RequestURI = ilu.CString; (* Should we handle URI parsing??? TYPE RequestURI = RECORD scheme : ilu.CString, net_loc : ilu.CString, path : ilu.CString, params : ilu.CString, query : ilu.CString, fragment : ilu.CString END; *) TYPE Header = RECORD name : field-name, value : optional-field-value END; TYPE HTTPHeader = Header; TYPE HTTPHeaders = SEQUENCE of HTTPHeader; TYPE EntityBody = SEQUENCE of BYTE; TYPE OptionalEntityBody = OPTIONAL EntityBody; TYPE Request = RECORD URI : RequestURI, headers : HTTPHeaders, body : OptionalEntityBody END; TYPE StatusCode = ENUMERATION OK = 200, Created = 201, Accepted = 202, NoContent = 204, MovedPermanently = 301, MovedTemporarily = 302, NotModified = 304, BadRequest = 400, Unauthorized = 401, Forbidden = 403, NotFound = 404, InternalError = 500, NotImplemented = 501, BadGateway = 502, ServiceUnavailable = 503 END; TYPE Response = RECORD status : StatusCode, headers : HTTPHeaders, body : OptionalEntityBody END; TYPE ConnectionParameter = Header; TYPE Connection = SEQUENCE of ConnectionParameter; TYPE Resource = OBJECT METHODS GET (request: Request, connection: Connection) : Response, HEAD (request: Request, connection: Connection) : Response, POST (request: Request, connection: Connection) : Response END; TYPE OptionalResource = OPTIONAL Resource;
(* $Id: WD-ilu-requestor-960307.html,v 1.4 1996/03/07 17:02:48 connolly Exp $ *) (* I've thought about just eliminating this ISL and using HTTP to do logging, but I'm sticking with this right now to allow logging to be asynchronous. Comments? *) (* The following list is are the name-value pairs that must be contained in the headers (the separate requesters may include their own unique headers, and various clients might send different headers which should be passed along here): "content-length" == the length in bytes of the returned data "content-type" == the mime type of the returned data "method" == the method of the request "remote-ip" == the IP address of the remote client "remote-name" == the name of the remote client, or the IP address if the name cannot be determined "status" == the status code of the response "uri" == the URI of the request *) INTERFACE logger IMPORTS ilu, http END; TYPE LoggerObject = OBJECT METHODS ASYNCHRONOUS LogRequest(params: http.HTTPHeaders) END;
(* $Id: WD-ilu-requestor-960307.html,v 1.4 1996/03/07 17:02:48 connolly Exp $ *) INTERFACE authorize IMPORTS http END; TYPE NameType = ilu.CString; TYPE GroupList = SEQUENCE OF ilu.CString; EXCEPTION AuthenticationFailed; EXCEPTION Forbidden; EXCEPTION AuthorizationRequired: ilu.CString; TYPE AuthorizationRecord = RECORD name: ilu.CString, groups: GroupList END; TYPE OptionalAuthorizationRecord = OPTIONAL AuthorizationRecord; TYPE Authenticator = OBJECT METHODS AuthenticateUser(name: NameType, password: ilu.CString): AuthorizationRecord RAISES AuthenticationFailed END END; TYPE Authorizer = OBJECT METHODS AuthorizeUser(authorization-record: OptionalAuthorizationRecord) RAISES Forbidden, AuthorizationRequired END END;