I'm building a distributed C++ application that needs to do lots of serialization and deserialization of simple data structures that's being passed between different processes and computers.
I'm not interested in serializing complex class hierarchies, but more of sending structures with a few simple members such as number, strings and data vectors. The data vectors can often be many megabytes large.
I'm worried that text/xml-based ways of doing it is too slow and I really don't want to write this myself since problems like string encoding and number endianess can make it way more complicated than it looks on the surface.
I've been looking a bit at protocol buffers and boost.serialize. According to the documents protocol buffers seems to care much about performance.
Boost seems somewhat more lightweight in the sense that you don't have an external language for specifying the data format which I find quite convenient for this particular project.
So my question comes down to this: does anyone know if the boost serialization is fast for the typical use case I described above?
Also if there are other libraries that might be right for this, I'd be happy to hear about them.
I would strongly suggest protocol buffers. They're incredibly simple to use, offer great performance, and take care of issues like endianness and backwards compatibility. To make it even more attractive, serialized data is language-independent thanks to numerous language implementations.
ACE and ACE TAO come to mind, but you might not like the size and scope of it.
http://www.cs.wustl.edu/~schmidt/ACE.html
Regarding your query about "fast" and boost. That is a subjective term and without knowing your requirements (throughput, etc) it is difficult to answer that for you. Not that I have any benchmarks for the boost stuff myself...
There are messaging layers you can use, but those are probably slower than boost. I'd say that you identified a good solution in boost, but I've only used ACE and other proprietary communications/messaging products.
My guess is that boost is fast enough. I have used it in previous projects to serialize data to and from disk, and its performance never even came up as an issue.
My answer here talks about serialization in general, which may be helpful to you beyond which serialization library you choose to use.
Having said that, it looks like you know most of the main trouble spots with serialization (endianess string encoding). You did leave out versioning and forwards/backwards compatibility. If time is not critical I recommend writing your own serialization code. It is an enlightening experience, and the lessons you learn are invaluable. Though I will warn you it will tend to make you hate XML based protocols for their bloatedness. :)
Whichever path you choose good luck with your project.
Also check out ONC-RPC (old SUN-RPC)
boost.serialization doesn't care about string encodings or endianness. You'll be similarly well off not using it if that matters to you.
You might want to look into ICE from ZeroC: http://www.zeroc.com/
It works similar to CORBA, except that it's entirely specced and defined by the company. The upside is that the implementations work as intended, since there aren't all that many. The downside is that if you're using a language they don't support, you're out of luck.
If you are only sending well defined defined data structures, then perhaps you should be looking at ASN.1 as an encoding methodology ?
There's also Thrift, which looks like an alpha project but is used and developed by Facebook, so it has a few users of it.
Or good old DCE, which was the standard MS decided to use for COM. Its now open-source, 20 years too late, but better than never.
Don't pre-emptively optimize. Measure first and optimize second.
Related
I'm fairly new at a company as embedded software engineer and I have to deal with networking quite a lot. I have to program in C++ all the time and using the code of my colleagues I noticed they use structs to build up packets to send over a network using TCP/UDP. For this, because the machines we are using are little-endian, everything has to be converted to big-endian. So the moment the data has to be sent they use the known functions as htons, ntohs, etc.
So I was looking up on how to make my code portable so the programmer does not have to care about the endianness of his machine and I stumbled upon the Endian Arithmetic types of the Boost library with which you, the programmer, can decide how your integer types are stored in memory. They seem very nice in the context I would use them but they are mentioned barely anywhere. Stackoverflow only has posts on how the endianness problem can not be handled in a proper way but these Arithmetic types seem like a nice portable solution to me. Why is it these types are barely mentioned anywhere and are never a proposed solution? Is there something wrong with them or are they just too new?
It depends on where. Of course, using them throughout your whole codebase is a design trade-off.
The endian types are not a zero-cost abstraction. Using them for anything else than passive storage can hurt performance. (Imagine doing number crunching on non-native endian numbers. The processor could be spending more time dancing around with high-order/low-order bytes and words as needed than on the actual calculations).
If your application is typical, you will have POD transmission buffers that do not contain many "live" data elements, and the sheer benefit of automatic endianness conversions under the hood could well be worth any overhead.
I know about Boost Endian types and would prefer them over repeatedly writing ntoh and friends in a heartbeat.
In reality, though I've written generic serialization helpers (where the use of ntoh/hton is isolated to a handful primitive functions), or used existing libraries (such as protobuf). So the irony is that, me too, I haven't use Boost Endian extensively for any production code.
Perhaps this pattern goes for many developers, explaining not a lot of examples around Boost Endian?
I'm looking into a mechanism for serialize data to be passed over a socket or shared-memory in a language-independent mechanism. I'm reluctant to use XML since this data is going to be very structured, and encoding/decoding speed is vital. Having a good C API that's liberally licensed is important, but ideally there should be support for a ton of other languages. I've looked at google's protocol buffers and ASN.1. Am I on the right track? Is there something better? Should I just implement my own packed structure and not look for some standard?
Given your requirements, I would go with Google Protocol Buffers. It sounds like it's ideally suited to your application.
You could consider XDR. It has an RFC. I've used it and never had any performance problems with it. It was used in ONC RPC and has an and comes with a tool called rpcgen. It is also easy to create a generator yourself when you just want to serialize data (which is what I ended up doing for portability reasons, took me half a day).
There is an open source C implementation, but it can already be in a system library, so you wouldn't need the sources.
ASN.1 always seemed a bit baroque to me, but depending on your actual needs might be more appropriate, since there are some limitations to XDR.
Just wanted to throw in ASN.1 into this mix. ASN.1 is a format standard, but there's libraries for most languages, and the C interface via asn1c is much cleaner than the C interface for protocol buffers.
JSON is really my favorite for this kind of stuff. I have no prior experience with binary stuff in it though. Please post your results if you are planning on using JSON!
Thrift is a binary format created by Facebook. Here's a comparison with google protocol buffers.
Check out Hessian
There is also Binary XML but it seems not stabilized yet. The article I link to gives a bunch of links which might be of interest.
Another option is SNAC/TLV which is used by AOL in it's Oscar/AIM protocol.
Also check out Muscle. While it does quite a bit, it serializes to a binary format.
Few Thing's you need to Consider
1. Storage
2. Encoding Style (1 byte 2 byte)
3. TLV standards
ASN.1 Parser is the good for binary represenations the best part is ASN.1 is a well-established technology that is widely used both within ITU-T and outside of it. The notation is supported by a number of software vendors.
Before saying anything I have to say that, albeit I'm an experienced programmer in Java, I'm rather new to C / C++ programming.
I have to save a binary file in a format that makes it accessible from different operating systems & platforms. It should be very efficient because I have to deal with a lot of data. What approaches should I investigate for that? What are the main advantages and disadvantages?
Currently I'm thinking about using the network notation (something like htonl that is available both under unix and windows ). Is there a better way?
Network order (big-endian) is something of a de facto standard. However, if your program will be used mostly on x86 (which is little-endian), you may want to stick with that for performance reasons (the protocol will still be usable on big-endian machines, but they will instead have the performance impact).
Besides htonl (which converts 32-bit values), there's also htons (16-bit), and bswap_64 (non-standard for 64-bit).
If you want a binary format, but you'd like to abstract away some of the details to ease serialization and deserialization, consider Protocol Buffers or Thrift. Protocol Buffers are updatable (you can add optional or repeated (0 or more) fields to the schema without breaking existing code); not sure about Thrift.
However, before premature optimization, consider whether parsing is really the bottleneck. If reading every line of the file will require a database query or calculation, you may be able to use a more readable format without any noticeable performance impact.
I think there are a couple of decent choices for this kind of task.
In most cases, my first choice would probably be Sun's (now Oracle's) XDR. It's used in Sun's implementation of RPC, so it's been pretty heavily tested for quite a while. It's defined in RFC 1832, so documentation is widely available. There are also libraries (portable and otherwise) that know how to convert to/from this format. On-wire representation is reasonably compact and conversion fairly efficient.
The big potential problem with XDR is that you do need to know what the data represents to decode it -- i.e., you have to (by some outside means) ensure that the sender and receiver agree on (for example) the definition of the structs they'll send over the wire, before the receiver can (easily) understand what's being sent.
If you need to create a stream that's entirely self-describing, so somebody can figure out what it contains based only on the content of the stream itself, then you might consider ASN.1. It's crufty and nasty in some ways, but it does produce self-describing streams, is publicly documented, and it's used pretty widely (though mostly in rather limited domains). There are a fair number of libraries that implement encoding and decoding. I doubt anybody really likes it much, but if you need what it does, it's probably the first choice, if only because it's already known and somewhat accepted.
My first choice for a situation like this would be ASN.1 since it gives you the flexibility of using whatever programing language you desire on either end, as well as being platform independent. It hides the endian-ness issues from you so you don't have to worry about them. One end can use Java while the other end uses C or C++ or C#. It also supports multiple encoding rules you can chose from depending on your needs. There is PER (Packed Encoding Rules) if the goal is making the encoding as small as possible, or there is E-XER (Extended XML Encoding Rules) if you prefer to exchange information using XML, or there is DER (Distingushed Encoding Rules) if your application involves digital signatures or certificates. ASN.1 is widely used in telephony, but also used in banking, automobiles, aviation, medical devices, and several other area. It is mature proven technology that has stood the test of time and continues to be added in new areas where communication between disparate machines and programming languages is needed.
An excellent resource where you can try ASN.1 free is http://asn1-playground.oss.com where you can play with some existing ASN.1 specifications, or try creating your own, and see what the various encoding rules produce.
There are some excellent books available as a free download from http://www.oss.com/asn1/resources/books-whitepapers-pubs/asn1-books.html where the first one is titled "ASN.1 — Communication Between Heterogeneous Systems".
Closed. This question is opinion-based. It is not currently accepting answers.
Want to improve this question? Update the question so it can be answered with facts and citations by editing this post.
Closed 5 years ago.
Improve this question
Does anyone with experience with these libraries have any comment on which one they preferred? Were there any performance differences or difficulties in using?
I've been using Boost Serialization for a long time and just dug into protocol buffers, and I think they don't have the exact same purpose. BS (didn't see that coming) saves your C++ objects to a stream, whereas PB is an interchange format that you read to/from.
PB's datamodel is way simpler: you get all kinds of ints and floats, strings, arrays, basic structure and that's pretty much it. BS allows you to directly save all of your objects in one step.
That means with BS you get more data on the wire but you don't have to rebuild all of your objects structure, whereas protocol buffers is more compact but there is more work to be done after reading the archive. As the name says, one is for protocols (language-agnostic, space efficient data passing), the other is for serialization (no-brainer objects saving).
So what is more important to you: speed/space efficiency or clean code?
I've played around a little with both systems, nothing serious, just some simple hackish stuff, but I felt that there's a real difference in how you're supposed to use the libraries.
With boost::serialization, you write your own structs/classes first, and then add the archiving methods, but you're still left with some pretty "slim" classes, that can be used as data members, inherited, whatever.
With protocol buffers, the amount of code generated for even a simple structure is pretty substantial, and the structs and code that's generated is more meant for operating on, and that you use protocol buffers' functionality to transport data to and from your own internal structures.
There are a couple of additional concerns with boost.serialization that I'll add to the mix. Caveat: I don't have any direct experience with protocol buffers beyond skimming the docs.
Note that while I think boost, and boost.serialization, is great at what it does, I have come to the conclusion that the default archive formats it comes with are not a great choice for a wire format.
It's important to distinguish between versions of your class (as mentioned in other answers, boost.serialization has some support for data versioning) and compatibility between different versions of the serialization library.
Newer versions of boost.serialization may not generate archives that older versions can deserialize. (the reverse is not true: newer versions are always intended to deserialize archives made by older versions). This has led to the following problems for us:
Both our client & server software create serialized objects that the other consumes, so we can only move to a newer boost.serialization if we upgrade both client and server in lockstep. (This is quite a challenge in an environment where you don't have full control of your clients).
Boost comes bundled as one big library with shared parts, and both the serialization code and the other parts of the boost library (e.g. shared_ptr) may be in use in the same file, I can't upgrade any parts of boost because I can't upgrade boost.serialization. I'm not sure if it's possible/safe/sane to attempt to link multiple versions of boost into a single executable, or if we have the budget/energy to refactor out bits that need to remain on an older version of boost into a separate executable (DLL in our case).
The old version of boost we're stuck on doesn't support the latest version of the compiler we use, so we're stuck on an old version of the compiler too.
Google seem to actually publish the protocol buffers wire format, and Wikipedia describes them as forwards-compatible, backwards-compatible (although I think Wikipedia is referring to data versioning rather than protocol buffer library versioning). Whilst neither of these is a guarantee of forwards-compatibility, it seems like a stronger indication to me.
In summary, I would prefer a well-known, published wire format like protocol buffers when I don't have the ability to upgrade client & server in lockstep.
Footnote: shameless plug for a related answer by me.
Boost Serialisation
is a library for writing data into a stream.
does not compress data.
does not support data versioning automatically.
supports STL containers.
properties of data written depend on streams chosen (e.g. endian, compressed).
Protocol Buffers
generates code from interface description (supports C++, Python and Java by default. C, C# and others by 3rd party).
optionally compresses data.
handles data versioning automatically.
handles endian swapping between platforms.
does not support STL containers.
Boost serialisation is a library for converting an object into a serialised stream of data. Protocol Buffers do the same thing, but also do other work for you (like versioning and endian swapping). Boost serialisation is simpler for "small simple tasks". Protocol Buffers are probably better for "larger infrastructure".
EDIT:24-11-10: Added "automatically" to BS versioning.
I have no experience with boost serialization, but I have used protocol buffers. I like protocol buffers a lot. Keep the following in mind (I say this with no knowledge of boost).
Protocol buffers are very efficient so I don't imagine that being a serious issue vs. boost.
Protocol buffers provide an intermediate representation that works with other languages (Python and Java... and more in the works). If you know you're only using C++, maybe boost is better, but the option to use other languages is nice.
Protocol buffers are more like data containers... there is no object oriented nature, such as inheritance. Think about the structure of what you want to serialize.
Protocol buffers are flexible because you can add "optional" fields. This basically means you can change the structure of protocol buffer without breaking compatibility.
Hope this helps.
boost.serialization just needs the C++ compiler and gives you some syntax sugar like
serialize_obj >> archive;
// ...
unserialize_obj << archive;
for saving and loading. If C++ is the only language you use you should give boost.serialization a serious shot.
I took a fast look at google protocol buffers. From what I see I'd say its not directly comparable to boost.serialization. You have to add a compiler for the .proto files to your toolchain and maintain the .proto files itself. The API doesn't integrate into C++ as boost.serialization does.
boost.serialization does the job its designed for very well: to serialize C++ objects :)
OTOH an query-API like google protocol buffers has gives you more flexibility.
Since I only used boost.serialization so far I cannot comment on performance comparison.
Correction to above (guess this is that answer) about Boost Serialization :
It DOES allow supporting data versioning.
If you need compression - use a compressed stream.
Can handle endian swapping between platforms as encoding can be text, binary or XML.
I never implemented anything using boost's library, but I found Google protobuff's to be more thought-out, and the code is much cleaner and easier to read. I would suggest having a look at the various languages you want to use it with and have a read through the code and the documentation and make up your mind.
The one difficulty I had with protobufs was they named a very commonly used function in their generated code GetMessage(), which of course conflicts with the Win32 GetMessage macro.
I would still highly recommend protobufs. They're very useful.
I know that this is an older question now, but I thought I'd throw my 2 pence in!
With boost you get the opportunity to I'm write some data validation in your classes; this is good because the data definition and the checks for validity are all in one place.
With GPB the best you can do is to put comments in the .proto file and hope against all hope that whoever is using it reads it, pays attention to it, and implements the validity checks themselves.
Needless to say this is unlikely and unreliable if your relying on someone else at the other end of a network stream to do this with the same vigour as oneself. Plus if the constraints on validity change, multiple code changes need to be planned, coordinated and done.
Thus I consider GPB to be inappropriate for developments where there is little opportunity to regularly meet and talk with all team members.
==EDIT==
The kind of thing I mean is this:
message Foo
{
int32 bearing = 1;
}
Now who's to say what the valid range of bearing is? We can have
message Foo
{
int32 bearing = 1; // Valid between 0 and 359
}
But that depends on someone else reading this and writing code for it. For example, if you edit it and the constraint becomes:
message Foo
{
int32 bearing = 1; // Valid between -180 and +180
}
you are completely dependent on everyone who has used this .proto updating their code. That is unreliable and expensive.
At least with Boost serialisation you're distributing a single C++ class, and that can have data validity checks built right into it. If those constraints change, then no one else need do any work other than making sure they're using the same version of the source code as you.
Alternative
There is an alternative: ASN.1. This is ancient, but has some really, really, handy things:
Foo ::= SEQUENCE
{
bearing INTEGER (0..359)
}
Note the constraint. So whenever anyone consumes this .asn file, generates code, they end up with code that will automatically check that bearing is somewhere between 0 and 359. If you update the .asn file,
Foo ::= SEQUENCE
{
bearing INTEGER (-180..180)
}
all they need to do is recompile. No other code changes are required.
You can also do:
bearingMin INTEGER ::= 0
bearingMax INTEGER ::= 360
Foo ::= SEQUENCE
{
bearing INTEGER (bearingMin..<bearingMax)
}
Note the <. And also in most tools the bearingMin and bearingMax can appear as constants in the generated code. That's extremely useful.
Constraints can be quite elaborate:
Garr ::= INTEGER (0..10 | 25..32)
Look at Chapter 13 in this PDF; it's amazing what you can do;
Arrays can be constrained too:
Bar ::= SEQUENCE (SIZE(1..5)) OF Foo
Sna ::= SEQUENCE (SIZE(5)) OF Foo
Fee ::= SEQUENCE
{
boo SEQUENCE (SIZE(1..<6)) OF INTEGER (-180<..<180)
}
ASN.1 is old fashioned, but still actively developed, widely used (your mobile phone uses it a lot), and far more flexible than most other serialisation technologies. About the only deficiency that I can see is that there is no decent code generator for Python. If you're using C/C++, C#, Java, ADA then you are well served by a mixture of free (C/C++, ADA) and commercial (C/C++, C#, JAVA) tools.
I especially like the wide choice of binary and text based wireformats. This makes it extremely convenient in some projects. The wireformat list currently includes:
BER (binary)
PER (binary, aligned and unaligned. This is ultra bit efficient. For example, and INTEGER constrained between 0 and 15 will take up only 4 bits on the wire)
OER
DER (another binary)
XML (also XER)
JSON (brand new, tool support is still developing)
plus others.
Note the last two? Yes, you can define data structures in ASN.1, generate code, and emit / consume messages in XML and JSON. Not bad for a technology that started off back in the 1980s.
Versioning is done differently to GPB. You can allow for extensions:
Foo ::= SEQUENCE
{
bearing INTEGER (-180..180),
...
}
This means that at a later date I can add to Foo, and older systems that have this version can still work (but can only access the bearing field).
I rate ASN.1 very highly. It can be a pain to deal with (tools might cost money, the generated code isn't necessarily beautiful, etc). But the constraints are a truly fantastic feature that has saved me a whole ton of heart ache time and time again. Makes developers whinge a lot when the encoders / decoders report that they've generated duff data.
Other links:
Good intro
Open source C/C++ compiler
Open source compiler, does ADA too AFAIK
Commercial, good
Commercial, good
Try it yourself online
Observations
To share data:
Code first approaches (e.g. Boost serialisation) restrict you to the original language (e.g. C++), or force you to do a lot of extra work in another language
Schema first is better, but
A lot of these leave big gaps in the sharing contract (i.e. no constraints). GPB is annoying in this regard, because it is otherwise very good.
Some have constraints (e.g. XSD, JSON), but suffer patchy tool support.
For example, Microsoft's xsd.exe actively ignores constraints in xsd files (MS's excuse is truly feeble). XSD is good (from the constraints point of view), but if you cannot trust the other guy to use a good XSD tool that enforces them for him/her then the worth of XSD is diminished
JSON validators are ok, but they do nothing to help you form the JSON in the first place, and aren't automatically called. There's no guarantee that someone sending you JSON message have run it through a validator. You have to remember to validate it yourself.
ASN.1 tools all seem to implement the constraints checking.
So for me, ASN.1 does it. It's the one that is least likely to result in someone else making a mistake, because it's the one with the right features and where the tools all seemingly endeavour to fully implement those features, and it is language neutral enough for most purposes.
To be honest, if GPB added a constraints mechanism that'd be the winner. XSD is close but the tools are almost universally rubbish. If there were decent code generators of other languages, JSON schema would be pretty good.
If GPB had constraints added (note: this would not change any of the wire formats), that'd be the one I'd recommend to everyone for almost every purpose. Though ASN.1's uPER is very useful for radio links.
As with almost everything in engineering, my answer is... "it depends."
Both are well tested, vetted technologies. Both will take your data and turn it into something friendly for sending someplace. Both will probably be fast enough, and if you're really counting a byte here or there, you're probably not going to be happy with either (let's face it both created packets will be a small fraction of XML or JSON).
For me, it really comes down to workflow and whether or not you need something other than C++ on the other end.
If you want to figure out your message contents first and you're building a system from scratch, use Protocol Buffers. You can think of the message in an abstract way and then auto-generate the code in whatever language you want (3rd party plugins are available for just about everything). Also, I find collaboration simplified with Protocol Buffers. I just send over a .proto file and then the other team has a clear idea of what data is being transfered. I also don't impose anything on them. If they want to use Java, go ahead!
If I already have built a class in C++ (and this has happened more often than not) and I want to send that data over the wire now, Boost Serialization obviously makes a ton of sense (especially where I already have a Boost dependency somewhere else).
You can use boost serialization in tight conjunction with your "real" domain objects, and serialize the complete object hierarchy (inheritance). Protobuf does not support inheritance, so you will have to use aggregation. People argue that Protobuf should be used for DTOs (data transfer objects), and not for core domain objects themselves. I have used both boost::serialization and protobuf. The Performance of boost::serialization should be taken into account, cereal might be an alternative.
Do you use Luabind, toLua++, or some other library (if so, which one) or none at all?
For each approach, what are the pro's and con's?
I can't really agree with the 'roll your own' vote, binding basic types and static C functions to Lua is trivial, yes, but the picture changes the moment you start dealing with tables and metatables; things go trickier very quickly.
LuaBind seems to do the job, but I have a philosophical issue with it. For me it seems like if your types are already complicated the fact that Luabind is heavily templated is not going to make your code any easier to follow, as a friend of mine said "you'll need Herb Shutter to figure out the compilation messages". Plus it depends on Boost, plus compilation times get a serious hit, etc.
After trying a few bindings, Tolua++ seems the best. Tolua doesn't seem to be very much in development, where as Tolua++ seems to work fine (plus half the 'Tolua' tutorials out there are, in fact, 'Tolua++' tutorials, trust me on that:) Tolua does generate the right stuff, the source can be modified and it seems to deal with complicated cases (like templates, unions, nameless structs, etc, etc)
The biggest issue with Tolua++ seems to be the lack of proper tutorials, pre-set Visual Studio projects, or the fact that the command line is a bit tricky to follow (you path/files can't have white spaces -in Windows at least- and so on) Still, for me it is the winner.
To answer my own question in part:
Luabind: once you know how to bind methods and classes via this awkward template syntax, it's pretty straightforward and easy to add new bindings. However, luabind has a significant performance impact and shouldn't be used for realtime applications. About 5-20 times more overhead than calling C functions that manipulate the stack directly.
I don't use any library. I have used SWIG to expose a C library some time ago, but there was too much overhead, and I stop using it.
The pros are better performance and more control, but its takes more time to write.
Use raw Lua API for your bindings -- and keep them simple. Take inspiration in the API itself (AUX library) and libraries by Lua authors.
With some practice raw API is the best option -- maximum flexibility and minimum of unneeded overhead. You've got what you want and no more, the way you need it to be.
If you must bind large third-party libraries use automated generators like tolua, tolua++ (or even roll your own for the specific case). It would free you from manual work.
I would not recommend using Luabind. At the moment it's development stalled (however starting to come back to life), and if you would meet some corner case, you may be on your own. Also Luabind heavily uses template metaprogramming. This may (and may not) be unacceptable, depending on the point of view.